Vue normale

Aujourd’hui — 18 juillet 2025Numérique
  • ✇404 Media
  • Behind the Blog: High Stakes Data Dumps
    This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss data dumps, high stakes, and lizard brain screen time.JOSEPH: Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador is the hardest hacking related article we’ve ever worked on. I’ve obtained some very sensitive data breaches over the last decade: metadata of specific individuals from the mas
     

Behind the Blog: High Stakes Data Dumps

18 juillet 2025 à 13:26
Behind the Blog: High Stakes Data Dumps

This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss data dumps, high stakes, and lizard brain screen time.

JOSEPH: Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador is the hardest hacking related article we’ve ever worked on. 

I’ve obtained some very sensitive data breaches over the last decade: metadata of specific individuals from the massive AT&T breach; photos of peoples’ genitalia pre- and post-plastic surgery. Honestly it’s hard to remember them all.

The data here wasn’t even necessarily as sensitive or personal as those. It was flight manifests, which contain peoples’ names, the flight they were on, and their gender. That’s basically it. But it was how to handle publication of the data that was exceptionally complicated and why it took us a while from when we first obtained the data a few months ago to publishing this week.

  • ✇404 Media
  • Come Party With Us In Los Angeles July 30
    Hello 404 Media readers! We're excited to announce that we're having our first ever party in Los Angeles. We have partnered with the amazing DIY hackerspace RIP.SPACE in the Arts District. We'll start the night with a live podcast about the surveillance technologies powering ICE, with a specific focus on tools that are being used in Los Angeles. We'll then change gears and do some Q&A about 404 Media and independent journalism. I'm considering doxing my Instagram algorithm as well. After tha
     

Come Party With Us In Los Angeles July 30

18 juillet 2025 à 12:52
Come Party With Us In Los Angeles July 30

Hello 404 Media readers! We're excited to announce that we're having our first ever party in Los Angeles. We have partnered with the amazing DIY hackerspace RIP.SPACE in the Arts District.

We'll start the night with a live podcast about the surveillance technologies powering ICE, with a specific focus on tools that are being used in Los Angeles. We'll then change gears and do some Q&A about 404 Media and independent journalism. I'm considering doxing my Instagram algorithm as well. After that, we'll have a reception and party with music from our friend DJ Avey.

We'll have free beer and wine, good vibes, and hopefully a good conversation. Tickets are free for subscribers, $10 for the general public (you can also subscribe for free entry here). If you're a subscriber, scroll to the end of this post for your free ticket code.

GET TICKETS HERE

.

.

.

.

.

  • ✇404 Media
  • White House Partners With PragerU to Make AI-Slopified Founding Fathers
    Conservative content mill PragerU is partnering with the White House to make AI-generated videos of founding fathers and Revolutionary War-era randos.PragerU is a nonprofit organization with a mission “to promote American values through the creative use of digital media, technology and edu-tainment,” according to its website. It’s been criticized for advancing climate denial and slavery apologism, frequently publishes videos critical of “wokeness” and “DEI,” and is very concerned about “the d
     

White House Partners With PragerU to Make AI-Slopified Founding Fathers

18 juillet 2025 à 10:09
White House Partners With PragerU to Make AI-Slopified Founding Fathers

Conservative content mill PragerU is partnering with the White House to make AI-generated videos of founding fathers and Revolutionary War-era randos.

PragerU is a nonprofit organization with a mission “to promote American values through the creative use of digital media, technology and edu-tainment,” according to its website. It’s been criticized for advancing climate denial and slavery apologism, frequently publishes videos critical of “wokeness” and “DEI,” and is very concerned about “the death of the West.” It has also been increasingly integrated into school curricula around the country.

PragerU held a launch event for the series, “Road to Liberty,” on June 25. Secretary Linda McMahon took some time away from dismantling the Department of Education to speak at the event. In person at the White House, visitors can tour a display of notable Revolutionary War people and places, and scan a QR code on displays that take them to PragerU’s AI-generated videos of people from that time period speaking. 

Each of the videos highlights a different person who was alive during the signing of the Declaration of Independence, from former presidents to relatively minor players in the fight for independence. The videos are clearly AI-generated, with the sepia-toned peoples’ mouths moving almost independently from the rest of their faces in some of them. In one, an AI-generated John Adams says “facts do not care about our feelings,” a phrase commonly attributed to conservative commentator and PragerU contributor Ben Shapiro. 

At the end of the videos, there's a logo for the White House with the text "brought to you by PragerU," and a disclaimer: "The White House is grateful for the partnership with PragerU and the U.S. Department of Education in the production of this museum. This partnership does not constitute or imply U.S. Government or U.S. Department of Education endorsement of PragerU."

Professor of history Seth Cotlar spotted the videos in a thread on Bluesky

I have the unfortunate duty to inform you that the WH & Dept of Ed, as part of the Trump Admin's celebration of the 250th anniversary of the Declaration of Independence, has partnered with Prager U to create AI-slop videos in which we see John Adams say "facts do not care about your feelings."

Seth Cotlar (@sethcotlar.bsky.social) 2025-07-16T14:47:30.314Z

I asked Cotlar, as someone who specializes in American history and the rise of the far-right, what stood out to him about these videos. I thought it was odd, I said, that they chose to include people like politician and disgraced minister Lyman Hall and obscure poet Francis Hopkinson alongside more well-known figures like John Adams or Thomas Jefferson. 

“You're right to note that it's a pretty odd collection of figures they've chosen,” Cotlar said. “My guess is that this is part of the broader right wing populist push to frame themselves as the grassroots ‘true Americans,’ and they're including all of these lesser known figures with the hopes that their viewers will be like ‘oh wow, look at all of these revolutionary freedom fighters like me who were just kinda ordinary guys like me but who still changed history.’” 

He also said it’s noteworthy that the “Road to Liberty” lineup so far is almost entirely white men, including the random dudes like Hall and Hopkinson. “The lack of any pretense to inclusion is pretty notable. Even conservative glosses on the Revolution from the pre-Trump era would have included things like the Rhode Island Regiment or Lemuel Haynes or Phyllis Wheatley. Needless to say, they absolutely do not include Deborah Sampson,” Cotlar said. All of the people in the “coming soon” section on PragerU’s website are also white men. 

AI slop has become the aesthetic of the right, with authoritarians around the world embracing ugly, lazy, mass-produced content like PragerU’s founding father puppets. Here in the U.S., we have President Donald Trump hawking it on his social media accounts, including AI-generated images of himself as the Pope and “Trump Gaza,” an AI video and song depicting the West Bank as a vacation paradise where Trump parties alongside his former bestie Elon Musk. As Republicans used the response to Hurricane Helene to blame migrants, Amy Kremer, founder of Women for Trump, posted an AI image of a child caught in a flood hugging a puppy and then said she didn’t care that it wasn’t real: “Y’all, I don’t know where this photo came from and honestly, it doesn’t matter,” she wrote on X. Mike Lee shared the same image. AI slop makes for quick and easy engagement farming, and now it’s being produced in direct partnership with the White House.

I’m not sure what app or program PragerU is using to make these videos. I thought, at first, that they might be using one of the many basic lipsyncing or “make this old photo come alive” mobile apps on the market now. But the videos look better, or at least more heavily produced, than most of those apps are capable of. Just to make sure they haven’t somehow advanced wildly in the last few months since I checked one out, I tried one of them, Revive, and uploaded an image of John Adams to see if it would return anything close to what PragerU’s putting out. It did not. 

0:00
/0:10

The PragerU videos aren't this bad, but they also aren’t as good as what would come out of Veo 3, the newest AI video generator, which generates highly realistic videos complete with sound and speech, from text prompts. I gave Veo a painting of John Adams and told it what to say; PragerU probably isn’t using this generator, because the result is much more realistic than what’s in the “Road to Liberty” series, even when I use a screenshot from one of their videos.

0:00
/0:08

John Adams in Veo 3 using a painting as a prompt.

On the off chance the culprit is Midjourney—although the series’ style and the way the subjects’ mouths move almost independently of the rest of their faces don’t match what I’ve seen of Midjourney’s videos—I tried that one, too. I just gave Midjourney the same Adams portrait and a prompt for it to animate him praising the United States and it returned a raving lunatic, silently screaming. 

0:00
/0:05

Striking out so far, I emailed Hany Farid, a professor at UC Berkeley and Chief Science Officer of synthetic media detection company GetReal, and asked if he had any leads. He said it looked similar to what comes out of AI video creation platform HeyGen, which creates AI talking heads and generates speech for them using ElevenLabs. I tried this on screenshots of the avatars in PragerU’s Martha Washington and John Adams videos to see if the puppet-mouth-style matched up, and they were pretty close.

0:00
/0:01

HeyGen John Adams

0:00
/0:06

HeyGen Martha Washington

PragerU’s videos are still more heavily produced than what I could make using the free version of HeyGen; it’s possible they used a combination of these to make the videos, plus some old-fashioned video editing and animation to create the final products. PragerU reported almost $70 million in income last year, they can afford the effort. 

“While the PragerU stuff is distinctly terrible, it's not like our culture has commemorated the Revolution with high-minded sophistication,” Cotlar told me. “I was 8 during the bicentennial and while I definitely learned some stuff about the founding era, most of what I absorbed was pretty schlocky.” He mentioned the "Bicentennial minutes" that were broadcast in 1975 and 76, sponsored by Shell, and which TV critic John J. O’Connor called “so insubstantial as to be almost meaningless.” The series won an Emmy.

In the last two years, several states, beginning with Florida, have approved PragerU content to be taught in public school classrooms. In Oklahoma, teachers relocating from states with "progressive education policies” will have to undergo an assessment in partnership with PragerU to determine if they’re allowed to teach. "If you want to teach here, you'd better know the Constitution, respect what makes America great, and understand basic biology,” State Superintendent Ryan Walters said in a press release. “We're raising a generation of patriots, not activists, and I'll fight tooth and nail to keep leftist propaganda out of our classrooms."

Hier — 17 juillet 2025Numérique

The Astronomer CEO's Coldplay Concert Fiasco Is Emblematic of Our Social Media Surveillance Dystopia

17 juillet 2025 à 14:56
Subscribe
Join the newsletter to get the latest updates.
Success
Great! Check your inbox and click the link.
Error
Please enter a valid email address.
The Astronomer CEO's Coldplay Concert Fiasco Is Emblematic of Our Social Media Surveillance Dystopia

The CEO seemingly having an affair with the head of HR at his company at the Coldplay concert is a viral video for the ages, but it is also, unfortunately, emblematic of our current private surveillance and social media hellscape.

The video, which is now viral on every platform that we can possibly think of, has been covered by various news outlets, and is Pop Crave official, shows Andy Byron, the CEO of a company called Astronomer, with his arms around Astronomer’s head of HR, Kristen Cabot. The jumbotron cuts from one fan to this seemingly happy couple. They both simultaneously die inside; “Oh look at this happy couple,” Coldplay lead singer Chris Martin says. The woman covers her face and spins away. The man ducks out of frame. “Either they’re having an affair or they’re very shy,” Martin said. The camera pans to another company executive standing next to them, who is seemingly shaking out of discomfort.

It is hard to describe how viral this is at the moment, in a world in which so many awful things are occurring and in which nothing holds anyone’s attention for any length of time and in a world in which we are all living in our own siloed realities. “Andy Byron” is currently the most popular trending Google term in the United States, with more than double the searches of the next closest term. 

There are so many levels to this embarrassment—the Coldplay of it all, the HR violation occurring on jumbotron, etc—that one could likely write a doctoral dissertation on this 15 second video.

0:00
/0:19
  • ✇404 Media
  • Inside ICE’s Supercharged Facial Recognition App of 200 Million Images
    ICE officers are able to point their smartphone’s camera at a person and near instantaneously run their face against a bank of 200 million images, then pull up their name, date of birth, nationality, unique identifiers such as their “alien” number, and whether an immigration judge has determined they should be deported from the country, according to ICE material viewed by 404 Media.The new material, which includes user manuals for ICE’s recently launched internal app called Mobile Fortify, pr
     

Inside ICE’s Supercharged Facial Recognition App of 200 Million Images

17 juillet 2025 à 12:31
Inside ICE’s Supercharged Facial Recognition App of 200 Million Images

ICE officers are able to point their smartphone’s camera at a person and near instantaneously run their face against a bank of 200 million images, then pull up their name, date of birth, nationality, unique identifiers such as their “alien” number, and whether an immigration judge has determined they should be deported from the country, according to ICE material viewed by 404 Media.

The new material, which includes user manuals for ICE’s recently launched internal app called Mobile Fortify, provides granular insight into exactly how ICE’s new facial recognition app works, what data it can return on a subject, and where ICE is sourcing that data. The app represents an unprecedented linking of government databases into a single tool, including from the State Department, Customs and Border Protection (CBP), the FBI, and state records. It also includes the potential for ICE to later add commercially available databases that contain even more personal data on people inside the United States.

💡
Do you know anything else about this app? I would love to hear from you. Using a non-work device, you can message me securely on Signal at joseph.404 or send me an email at joseph@404media.co.

“This app shows that biometric technology has moved well beyond just confirming someone's identity. In the hands of ICE officers, it's becoming a way to retrieve vast amounts of data about a person on demand just by pointing a camera in their face,” Dave Maass, director of investigations at the Electronic Frontier Foundation (EFF), told 404 Media. “The more they streamline its use, the more they streamline its abuse. When an officer says, ‘papers please,’ you could choose to say nothing and face the consequences; with face recognition, your options are diminished.”

  • ✇404 Media
  • Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador
    The flight manifests for three legally contested deportation flights from Texas to El Salvador contain dozens of additional, unaccounted for passengers than a previously published Department of Homeland Security (DHS) list of people deported from the United States on those flights, 404 Media has learned. The additional people on the flight manifest have not been publicly acknowledged by the U.S. government in any way, and immigration experts who have been closely monitoring Trump’s deportatio
     

Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador

17 juillet 2025 à 11:52
Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador

The flight manifests for three legally contested deportation flights from Texas to El Salvador contain dozens of additional, unaccounted for passengers than a previously published Department of Homeland Security (DHS) list of people deported from the United States on those flights, 404 Media has learned. The additional people on the flight manifest have not been publicly acknowledged by the U.S. government in any way, and immigration experts who have been closely monitoring Trump’s deportation campaign say they have no idea where these people are or what happened to them. 404 Media is now publishing the names of these people. 

On March 15, the Trump administration deported more than 200 people on three aircraft to a megaprison in El Salvador. A judge blocked the deportations, but hours later the flights still landed in the country. It marked one of the major turning points of the administration’s mass deportation efforts, and signaled what was to come around the country—a lack of due process, authorities ignoring judge’s rulings, and deporting people on the flimsiest of pretenses. Soon after these flights, CBS News published an “internal government list” of people it said were deported to CECOT, the notorious El Salvadorian megaprison.

But in May, a hacker targeted GlobalX, the airline that operated these flights and shared the data with 404 Media. In addition to the names of people who were on the list CBS News published, the GlobalX flight manifests contain the names of dozens of people who were supposedly on the flights but whose status and existence has not been acknowledged by the U.S. government or previously reported in the press. 

“We have this list of people that the U.S. government has not formally acknowledged in any real way and we pretty much have no idea if they are in CECOT or someplace else, or whether they received due process,” Michelle Brané, executive director of Together and Free, a group that has been working with families of deported people, told 404 Media. “I think this further demonstrates the callousness and lack of due process involved and is further evidence that the US government is disappearing people. These people were detained and no one knows where they are, and we don't know the circumstances […] For almost all of these people, there’s no records whatsoever. No court records, nothing.” 

💡
Do you know anything else about these people or flights? We would love to hear from you. Using a non-work device, you can message Jason securely on Signal at jason.404 or send an email to jason@404media.co. You can Signal Joseph at joseph.404 or email joseph@404media.co.

“[The government is] not disclosing it and they’ve presumably been sent to a prison or sent somewhere by the U.S. government on a plane and have never been heard from since,” she added. “We have not heard from these people’s families, so I think perhaps even they don’t know.”

Brané added that it remains entirely unclear whether all of these people were actually on the flights or why they were on the manifests. If they were indeed on the flights, it is unknown where they currently are. That uncertainty, and the unwillingness of the U.S. government to provide any clarity about these people, is a major problem, she said.

While the stories of some of the people deported on these flights have garnered a lot of attention, such as Kilmar Abrego Garcia, U.S. authorities have refused to reveal the names of everyone on board. 

While the whereabouts and circumstances of most of these people remain unknown, Brané’s organization used publicly available data to try to better understand who they are. In some cases, Together and Free was able to identify a few details about specific people on the manifest. For example, one person on the manifest appears to have been arrested by local police in Texas in late December on drug possession charges and is listed in arrest records as being an “illegal alien.” Another person was arrested in Nashville in February on charges of driving without a license. For many other people listed, there is no easily discernible public data about who they are or why they appeared on the flight manifest.

Several other people are on the flight manifests and do not appear on the CBS News list, but their identities had already become public because their families have filed lawsuits or have been looking for them on social media. These include Abrego Garcia and Ricardo Prada Vásquez, a man whose family said he was “disappeared” because he did not appear on any official, publicly published lists. After the New York Times published an article about his disappearance, the Trump administration said he was at CECOT, and 404 Media was able to find his name on the March 15 flight manifests. 

In Venezuela, the family of another man who appears on the flight manifests but not on the CBS News list, Keider Alexander Flores Navas, has been protesting his disappearance and demanding answers. In a TikTok video posted in March, his mother Ana Navas explains that they suddenly stopped hearing from Keider before the March 15 flights. She said she eventually heard he was in federal detention. Then, she saw a photo of him in CECOT amongst a group of other prisoners: “The thing that worried me the most was he was not on any list. But this photo is from El Salvador. Lots of family members here recognize their sons [in official CECOT photos]. That’s my son,” she says, the camera panning to a circled image of Keider in CECOT.

Flight Manifests Reveal Dozens of Previously Unknown People on Three Deportation Flights to El Salvador

In another TikTok video posted in June, the mother of 21-year-old Brandon Sigaran-Cruz explains that he had been “disappeared for three months” with no news of his whereabouts. Sigaran-Cruz also appears on the flight manifest but not the CBS News list. 

 The U.S. government previously acknowledged that, along with more than 200 Venezuelan citizens, it deported 23 Salvadorans to El Salvador on the three March 15 flights. There is no formal list of the Salvadorans who were on the flight, and none of them appeared on the CBS News list, which included only Venezuelan citizens. 

The United Nations’ Human Rights Office has also filed court petitions saying that it is investigating the “involuntary disappearances” of at least four Venezuelans who were sent to El Salvador on these flights. “Neither the Government of El Salvador nor the Government of the United States has published official information on the list of deported persons or their current place of detention,” the United Nations said in a “Report on Enforced or Involuntary Disappearances” it filed in court.

“There continues to be very little clarity as to the fate and whereabouts of the Venezuelans removed to El Salvador. To date, no official lists of the deported detainees have been published. Provision of further information by authorities is key, including providing families and their counsel with available information on the specific situation and whereabouts of their loved ones,” Elizabeth Throssell, a spokesperson for the UN Human Rights Office, told 404 Media in an email. “The UN Human Rights Office has been in contact with family members of over 100 Venezuelans believed to have been deported to El Salvador.”

404 Media asked the Department of Homeland Security (DHS) over multiple weeks if the agency had any legitimate security concerns with these names being published, or if it could tell us anything about these people. The agency never responded, despite responding to requests for comment for other 404 Media articles. GlobalX did not respond to a request for comment either.

“It is critical that we know who was on these March 15 flights,” Lee Gelernt, a lawyer at the American Civil Liberties Union (ACLU) and the lead counsel on the ACLU’s related case, told 404 Media. “These individuals were sent to a gulag-type prison without any due process, possibly for the remainder of their lives, yet the government has provided no meaningful information about them, much less the evidence against them. Transparency at a time like this is essential.”

In recent months, the U.S. government has said that the El Salvadorian government has jurisdiction over the people detained in CECOT, while El Salvador told the United Nations that “the jurisdiction and legal responsibility for these persons lie exclusively with the competent foreign authorities [the United States],” leading to a situation where people are detained in a foreign prison but both responsible parties are not willing to claim legal responsibility for them. A similar situation has happened in Florida at the “Alligator Alcatraz” camp, where people detained by the federal government are being held in a state-run facility, and experts have said it’s not clear who is in charge. Brané said with the massive increase in ICE funding as part of Trump’s new law, we are likely to see more detention camps, more detainments, more deportation flights, and, likely, more people who aren’t publicly accounted for in any way.

“When you look at what ICE is doing now in terms of how they treat people, how they operate when they're given even a little bit of rope, it’s terrifying to think what the budget increase is going to do,” Brané said. “This is a taste of what we're going to see on a much larger scale."

You can read the list below. 404 Media has removed people listed on the flight manifests as “guards” (404 Media found at least one of these names matched someone who lists their employment online as a flight transport detention officer). Reportedly eight women deported to El Salvador were later returned. 404 Media is not publishing the names of women known to have returned to the U.S. The manifest also includes the names of several El Salvadorians mentioned as being deported in a White House Press release, court proceedings, and media reports. We have not included their names below because the administration has formally acknowledged that they were deported.

Manuel Quijada-Leon
Irvin Quintanilla-Garcia
Jose Ramirez-Iraheta
Josue Rivera-Portillo
Jorge Rodriguez Gomez
Mario Jeavanni Rojas
Edgar Leonel Sanchez Rosales
Brandon Sigaran-Cruz
Miguel Enriquez Saravia
Abraham Hernandez-Mania
Jean Morales-Loaiza
Nelson Alfaro-Orellana
Jhonnarty Pachecho-Chirinos
Cristian Alpe-Tepas
Jordyn Alexander Alvarez
Jose Alvarez Gonzalez
Wilfredo Avendano Carrizalez
Jose Gregorio Buenano Cantillo
Istmar Campos Mejia
Jose Chanta-Ochoa
Keider Alexander Flores Navas
Noe Florez-Valladares
Miguel Fuentes-Lopez
Roberto Interiano Uceda
Jose Lopez Cruz
Diego Maldonado-Fuentes
William Martinez-Ruano
Osmer Mejias-Ruiz
Iran Ochoa Suescun
David Orantez Gonzalez
Ariadny Araque-Cerrada
Elena Cuenca Palma
Maria Franco Pina
Mayerkis Guariman Gonzalez
Wilmary Linares-Marcano
Scarlet Mendoza Perez
Ofreilimar Peña Boraure
Edilianny Stephany Rivero Sierralta
Dioneli Sanz Aljorna
Anyeli Sequera Ramirez
Yanny Suarez Rodriguez
Karla Villasmil-Castellano

  • ✇404 Media
  • The 800-Year-Old Mystery of a Lost Medieval Legend Is Solved, Researchers Say
    🌘Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week. A major mystery about a long-lost legend that was all the rage in Medieval England but survives in only one known fragment has been solved, according to a study published on Tuesday in The Review of English Studies.Roughly 800 years ago, a legend known today as the Song of Wade was a blockbuster hit for English audiences. Mentions of the heroic character sho
     

The 800-Year-Old Mystery of a Lost Medieval Legend Is Solved, Researchers Say

17 juillet 2025 à 10:11
🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
The 800-Year-Old Mystery of a Lost Medieval Legend Is Solved, Researchers Say

A major mystery about a long-lost legend that was all the rage in Medieval England but survives in only one known fragment has been solved, according to a study published on Tuesday in The Review of English Studies.

Roughly 800 years ago, a legend known today as the Song of Wade was a blockbuster hit for English audiences. Mentions of the heroic character showed up in the works of Geoffrey Chaucer, for example. But the tale vanished from the literature centuries later, puzzling generations of scholars who have tried to track down its origin and intent. 

Now, for the first time, researchers say they’ve deciphered its true meaning—which flies in the face of the existing interpretation. 

“It is one of these really interesting and very unusual situations where we have a legend that was widely known and hugely popular throughout the Middle Ages, and then very suddenly in the middle of the 16th century, in the High Renaissance, it's just completely lost,” said James Wade, fellow in English at Girton College, University of Cambridge, who co-authored the study.

In 1896, the Medieval scholar M.R. James made a breakthrough on this literary cold case when he discovered a fragment of the Song of Wade in the Humiliamini sermon, which is part of a compendium that dates back to the 12th century. James brought the text to his colleague Israel Gollancz, a philologist with expertise in early English literature, and together they worked on a translation.

It is “the only surviving fragment” of the Song of Wade, said co-author Seb Falk, fellow in history and philosophy of science also at Girton College. “Obviously it had been around already for a while by the time this was written because it's part of the culture. The person who writes this sermon clearly expects his listeners to understand it and to know what he's talking about.”

James and Gollancz “knew there was no surviving text and they understood what they were looking at,” added Wade. “This was big news. It made the papers in 1896.”

But while the 19th century scholars recovered the sermon, their translation only deepened the mystery of the enigmatic text. For instance, references to “elves” and “sprites” in the translation suggest that the Song of Wade falls into a genre of fantastical epics about supernatural monsters. But when Chaucer references Wade in his works Troilus and Criseyde and The Merchant’s Tale, he places the character in a totally different tradition of chivalric romances which are rich with metaphors, but typically favor more grounded scenarios.

Chaucer’s mentions of Wade have perplexed scholars for centuries: in 1598, for instance, an early Chaucer editor named Thomas Speght wrote: “Concerning Wade and his [boat] called Guingelot, as also his strange exploits in the same, because the matter is long and fabulous, I passe it over.” In other words, Speght didn’t even try to decipher what Chaucer meant with his references to Wade.

This punt has become legendary in literary circles. “F. N. Robinson wrote in 1933 that Speght’s comment ‘has often been called the most exasperating note ever written on Chaucer’, and Richard Firth Green observed that Speght’s note ‘has caused generations of scholars to tear their hair out,’” Wade and Falk write in their new study.  

In 1936, the scholar Jack Bennett "supposed that there is ‘probably no better known crux in Chaucer than the tale of Wade,’” the pair added.

A few years ago, Wade and Falk set out to see if they could shed light on this famous and persistent riddle. Like so many good ideas, it began over a lunchtime conversation. From there, the team slowly and methodically worked through the sermon, scrutinizing each letter and rune.  

“Just trying to decipher the thing took quite a lot of work—transcribing it and then making an initial translation,” Falk said. “But I started realizing there's some really interesting material here from my point of view, as a historian of science, with lots of animals mentioned, both in Middle English and in Latin.”

As the pair worked through the text, they began to suspect that the scribe who originally copied the work may not have been very familiar with Middle English, leading to some transcription errors with certain runes. In particular, they found that the longstanding translation of “elves” and “sprites” were, in their view, more likely to be “wolves” and “sea snakes.” This transforms a key Song of Wade passage, “Some are elves and some are adders; some are sprites that dwell by waters,” to: 

“Some are wolves and some are adders; some are sea-snakes that dwell by the water.”

It may seem like a subtle shift, but it is a major sea change for the interpretation of the text. The switch to animals, as opposed to supernatural beings, suggests that the preacher who wrote the sermon was using animals as metaphors for human vices and behaviors—a reading that provides a much better fit for the overall sermon, and at last explains why Chaucer viewed Wade in the tradition of chivalry.

“It became pretty clear to us that the scribe had probably made some kind of a mistake because he was used to writing Latin rather than English,” Falk said. “We're not here to say that people who read it differently previously were stupid to miss it, but I think by looking at it from the outside in, we got a different perspective.”

“It then radically changes the meaning of the passage from being about monsters to being about animals, and therefore, changes it from being a piece about mythical beasts to a piece about courtly romance,” he added.

In addition to relieving centuries-old headaches over the legend, Wade and Falk also speculate that the sermon was originally written by English poet and abbot Alexander Neckam (1157–1217), based on its style and context clues. 

For the researchers, the thrill of the discovery lies not only in decoding the fragment, but in recovering a missing piece of cultural memory. While their new translation and attribution to Neckam are both tentative and may be disputed by other scholars, the study still opens a window into a long-lost legend and shows how fresh eyes can uncover insights in even the most perplexing fragments.

“By putting a completely different slant on it and, we think, understanding it properly, we have come much closer to the true meaning of the Wade legend,” Falk said. “Now, obviously we've only got three lines of this presumably much longer poem, and therefore, we can't pretend that we understand the thing in full, but I think we can understand it much better than we ever have before.”

“It reminds us that as time moves on, there's always the possibility of generational amnesia, of forgetting things, of losing things, and that when you have a chance to get a little bit back of something that humanity has lost, or that culture has lost, it's a really exciting moment,” Wade concluded.

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
À partir d’avant-hierNumérique
  • ✇404 Media
  • Steam Bends to Payment Processors on Porn Games
    Steam, the dominant digital storefront for PC games operated by Valve, updated its guidelines to forbid “certain kinds of adult content” and blamed restrictions from payment processors and financial institutions. The update was initially spotted by SteamDB.info , a platform that tracks and publishes data about Steam, and reported by the Japanese gaming site Gamespark.The update is yet another signal that payment processors are lately becoming more vigilant about what online platforms that hos
     

Steam Bends to Payment Processors on Porn Games

16 juillet 2025 à 11:44
Steam Bends to Payment Processors on Porn Games

Steam, the dominant digital storefront for PC games operated by Valve, updated its guidelines to forbid “certain kinds of adult content” and blamed restrictions from payment processors and financial institutions. The update was initially spotted by SteamDB.info , a platform that tracks and publishes data about Steam, and reported by the Japanese gaming site Gamespark.

The update is yet another signal that payment processors are lately becoming more vigilant about what online platforms that host adult content they’ll provide services to and another clear sign that they are currently the ultimate arbiter of what kind of content can be made easily available online, or not. 

Steam’s policy change appears under the onboarding portion of its Steamworks documentation for developers and publishers. The 15th item on a list of “what you shouldn’t publish on Steam” now reads: “Content that may violate the rules and standards set forth by Steam’s payment processors and related card networks and banks, or internet network providers. In particular, certain kinds of adult only content.”

It’s not clear when exactly Valve updated this list, but an archive of this page from April shows that it only had 14 items then. Other items that were already on the list included “nude or sexually explicit images of real people” and “adult content that isn’t appropriately labeled and age-gated,” but Valve did not previously mention payment processors specifically. 

"We were recently notified that certain games on Steam may violate the rules and standards set forth by our payment processors and their related card networks and banks," Valve spokesperson Kaci Aitchison Boyle told me in an email. "As a result, we are retiring those games from being sold on the Steam Store, because loss of payment methods would prevent customers from being able to purchase other titles and game content on Steam. We are directly notifying developers of these games, and issuing app credits should they have another game they’d like to distribute on Steam in the future."

Valve did not respond to questions about where developers might find more details about payment processors’ rules and standards. 

SteamDB.info, which also tracks when games are added or removed from Steam, noted many adult games have been removed from Steam in the last 24 hours. Sex games, many of which are of very low quality and sometimes include very extreme content, have been common on Steam for years. In April, I wrote about a “rape and incest” game called No Mercy which the developers eventually voluntarily removed from Steam after pressure from users, media, and lawmakers in the UK. The majority of games I saw that were removed from Steam recently revolve around similar themes, but we don’t know if they were removed by the developers or Valve, and if they were removed by Valve because of the recent policy change. Games are removed from Steam every day for a variety of reasons, including expired licensing deals or developers no longer wanting to support a game. 

However, Steam’s policy change comes at a time that we’ve seen increased pressure from payment processors around adult content. We recently reported that payment processors have forced two major AI models sharing platforms, Civitai and Tensor.Art, to remove certain adult content.

Update: This story has been updated with comment from Valve. 

  • ✇404 Media
  • Podcast: The AI Exodus Begins
    We start this week with a series of articles from Emanuel about a crackdown in the AI industry. After the break, Sam tells us about the ‘Save Our Signs’ campaign which hopes to preserve the history of national parks. In the subscribers-only section, Jason rants about how AI will not save the media industry.  Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a pa
     

Podcast: The AI Exodus Begins

16 juillet 2025 à 11:27
Podcast: The AI Exodus Begins

We start this week with a series of articles from Emanuel about a crackdown in the AI industry. After the break, Sam tells us about the ‘Save Our Signs’ campaign which hopes to preserve the history of national parks. In the subscribers-only section, Jason rants about how AI will not save the media industry. 

Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a paid subscriber, check your inbox for an email from our podcast host Transistor for a link to the subscribers-only version! You can also add that subscribers feed to your podcast app of choice and never miss an episode that way. The email should also contain the subscribers-only unlisted YouTube link for the extended video version too. It will also be in the show notes in your podcast player.

  • ✇404 Media
  • The Real Future of AI Is Ordering Mid Chicken at Bojangles
    Yesterday I ordered my lunch from an AI operating a drive-thru. It was fine. Banal. Boring even. A new experience that I think will become routine in the future.The AI drive-thru operator isn’t cutting edge tech deployed in an upscale market to win over high value consumers. I live at the edge of a South Carolina city with a little more than 140,000 people. A booming metropolis with the best and the finest, it is not.There’s a lot of local fast food fried chicken joints here and one of them i
     

The Real Future of AI Is Ordering Mid Chicken at Bojangles

16 juillet 2025 à 09:19
The Real Future of AI Is Ordering Mid Chicken at Bojangles

Yesterday I ordered my lunch from an AI operating a drive-thru. It was fine. Banal. Boring even. A new experience that I think will become routine in the future.

The AI drive-thru operator isn’t cutting edge tech deployed in an upscale market to win over high value consumers. I live at the edge of a South Carolina city with a little more than 140,000 people. A booming metropolis with the best and the finest, it is not.

There’s a lot of local fast food fried chicken joints here and one of them is Bojangles. It’s mid. Better than KFC and not as good as Popeyes, Bojangles is fine if you’re hungry but you’ll forget the meal as soon as it’s done and you’ll never yearn for it. Last year the restaurant said it would deploy an AI agent at its drive-thru windows. It’s called, I shit you not, Bo-Linda and made by the Israeli tech firm Hi-Auto.

According to the Bojangles website, “Bo-Linda™ can take guest orders 96+% of the time with no human intervention,” and “improve overall satisfaction by offloading order taking from team members and providing a consistent guest experience.”

When Bo-Linda finally arrived in South Carolina, I went to see what the fuss was about. It was crushingly dull. A preview of a time in the near future, I think, when the AI bubble retracts and the agents are common. It took my order with an efficiency that, I’ll be honest, is not typical of the typical fast food worker. The worst part was its constant attempts to up-sell me.

0:00
/0:39

“Do you want to upgrade your drink to our new water-melon iced tea?” It asked.

“No thank you.”

“Would you like to add our new peach cobbler for $1.99?”

“No thank you.”

“May I get you anything else?”

“No, that’s it.”“Would you like to round up for military scholarships?”“No thank you.”

“You’re welcome. Thank you. Your total is $10.89.”

When 404 Media founder Joseph Cox watched the video of my interactions, he made fun of my “no thank yous.” What can I say? There’s an ingrained and often stifling politeness that’s bred into us in the American South. Even though I knew I was talking to a machine, I couldn’t not be nice to it.

My thought in the immediate aftermath is that the whole thing was painless. My order wasn’t complicated, but it was correct. The machine never stumbled over itself or asked for clarification. It knew what I wanted and the humans at the window gave it to me. A few conversations with friends and a quick scan of social media in the area show that other people have had much the same interactions with Bo-Linda.

The drive-thru AI, much like the chicken it sold me, is fine. Forgettable.

It was later, sitting at home, and doing a little research for the story that concerns popped up. OpenAI CEO Sam Altman has said that saying “please” and “thank you” to ChatGPT has cost the company tens of millions of dollars. How much water and energy had I burned being polite to Bo-Linda the chatbot?

Sometimes it feels like the answers to these questions don’t matter. We’re barreling forward into the AI future, whether we like it or not. Data centers are springing up across America and nuclear power plants are coming back online, so Bojangles can make a little more money and so people in the drive-thru can feel a little less friction before eating their meal.

This is how a new technology takes over, what it feels like right before it becomes ubiquitous. One day you wake up and the cameras are everywhere, able to recognize your face and chart your movements across the city you live in. One day you look up and everyone has their face buried in their phone. It happened by degrees, but so gradually you didn’t notice. There were signs along the way, dangers and warnings.

But mostly, it was fine, as boring and routine as ordering chicken at a drive-thru.

  • ✇404 Media
  • 3D Printing Patterns Might Make Ghost Guns More Traceable Than We Thought
    So-called 3D-printed ghost guns are untraceable firearms that can be assembled at home. But cutting edge work from a forensic expert in California and researchers at the University of Oklahoma may soon show investigators can trace a 3D printed object to the specific printer that made it.Weapons manufactured using 3D printers have been a subject of Biden-era legislation and recent Supreme Court scrutiny. It’s possible to download the blueprints for a firearm and build it in your home. There’s
     

3D Printing Patterns Might Make Ghost Guns More Traceable Than We Thought

16 juillet 2025 à 09:14
3D Printing Patterns Might Make Ghost Guns More Traceable Than We Thought

So-called 3D-printed ghost guns are untraceable firearms that can be assembled at home. But cutting edge work from a forensic expert in California and researchers at the University of Oklahoma may soon show investigators can trace a 3D printed object to the specific printer that made it.

Weapons manufactured using 3D printers have been a subject of Biden-era legislation and recent Supreme Court scrutiny. It’s possible to download the blueprints for a firearm and build it in your home. There’s no serial number to track and no store to scrutinize your purchase. Luigi Mangione used a ghost gun to allegedly assassinate United Healthcare CEO Brian Thompson.

The Hyperpersonalized AI Slop Silo Machine Is Here

15 juillet 2025 à 14:30
Subscribe
Join the newsletter to get the latest updates.
Success
Great! Check your inbox and click the link.
Error
Please enter a valid email address.
The Hyperpersonalized AI Slop Silo Machine Is Here

For a while, I have said that the AI slop endgame, for social media companies, is creating a hyper personalized feed full of highly specific content about anything one could possibly imagine. Because AI slop is so easy to make and because social media algorithms are so personalized, this means that Facebook, Instagram, TikTok, or YouTube can feed you anything they perceive its users to possibly want. So this means that AI slop makers are exploring ever more niche areas of content. 

Case in point: Facebook AI slop about the horrific and deadly Texas flood. Topical AI content about disasters, war, current events, and news stories are at this point so commonplace that they are now sadly barely notable, and AI-powered “misinformation” about horrible events are all over every social media feed I can think of. But as we document our descent into this hellhole, I thought some AI slop surfaced on Bluesky by Christina Stephens was particularly notable:

The Hyperpersonalized AI Slop Silo Machine Is Here
  • ✇404 Media
  • Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
    Yesterday, Silicon Valley trade publication The Information launched TITV, a live-streaming news program. For its first episode, the company’s founder Jessica Lessin interviewed Mark Zuckerberg about his recent attempts to hire AI executives away from other companies for huge sums of money.It is impossible to say what actually happened in the interview, however, because there were audio problems with the livestream—the Zuckerberg segment was completely silent, giving us a minute and 50 second
     

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

15 juillet 2025 à 11:40
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

Yesterday, Silicon Valley trade publication The Information launched TITV, a live-streaming news program. For its first episode, the company’s founder Jessica Lessin interviewed Mark Zuckerberg about his recent attempts to hire AI executives away from other companies for huge sums of money.

It is impossible to say what actually happened in the interview, however, because there were audio problems with the livestream—the Zuckerberg segment was completely silent, giving us a minute and 50 seconds of video of Zuckerberg’s face moving with no sound coming out.

In recent weeks, however, Zuckerberg has really ramped up his rhetoric about “superintelligence,” throwing hundreds of millions of dollars at AI researchers in attempts to hire them at Meta, and beginning work on massive data centers around the country to power his AI. Zuckerberg and Meta have talked endlessly about AI as a game changer, while being unable or unwilling to keep AI-generated misinformation and slop off of Meta platforms.

Meanwhile, there has been increased discussion about the possible power of using AI for journalism, which we have written about extensively. I saw the silent interview with Zuckerberg as both an opportunity to check out the awesome abilities of Meta’s AI, and to test AI tools at a task some of them are supposedly good at—reading lips. If these technologies worked as well as their creators have promised, it might be possible to provide the world with information from The Lost Mark Zuckerberg Interview.

Here is how my experiment went.

Because Zuckerberg was obviously talking about Meta AI, I thought I would try his tool first. I took video from the interview and asked Meta AI to lip read it to tell me what they were saying:

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

“I’d love to help! However, I’m a text-based AI and don’t have capability to visually lip read videos,” Meta AI said. It then suggested that I “find a professional lip reader or transcription service that specializes in lip reading videos” or “use automated lip reading software.” 

I then tried to upload still images to see if Meta AI could parse them:

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

When I uploaded additional images, Meta AI was also not able to provide any information about what was being said.

I then went to ChatGPT, because Zuckerberg is reportedly offering pay packages of up to $300 million to OpenAI staffers to come work at Meta. I uploaded the 1:50 video and ChatGPT told me “the video processing took too long and timed out.” I then uploaded a 25 second clip and it told me “the system is still timing out while trying to extract frames.” I then asked it to do the first five seconds and it said “even with the shorter clip and smaller scope (first 5 seconds), the system timed out.” I then asked for it to extract one single frame, and it said “it looks like the system is currently unable to extract even a single frame from the video file.” ChatGPT then asked me to take a screenshot of Zuckerberg. I sent it this: 

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

And ChatGPT said “the person appears to be producing a sound like ‘f’ or ‘v’ (as in ‘video’ or ‘very’),” but that “possibly ‘m’ or ‘b,’ depending on the next motion.” I then shared the 10 frames around that single screenshot, and ChatGPT said “after closely analyzing the progression of lip shapes and facial motion,” the “probable lip-read phrase” was “This is version.” I then uploaded 10 more frames and it said the “full phrase so far (high confidence): ‘This version is just.’”

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

I then decided to try to extract every frame from the video and upload it to ChatGPT.

I went to a website called frame-extractor.com and cut the video into 3,000 frames. After it had processed 700 of them, I tried to upload them to ChatGPT and it did not work. I then decided I would go 10 frames at a time from the beginning of the clip. Even though I sent an entirely different portion of the video and told ChatGPT we were starting from a different part of the video, it still said that the beginning of the video said “this version is.” I continued uploading frames, 10 at a time. These frames included both Lessin and Zuckerberg, not just Zuckerberg.

ChatGPT slowly began to create a surely accurate transcript of the lost audio of this interview: “This version is just that it we built,” ChatGPT said. As I added more and more frames, it refined the answer: “This version is what we’re going to do,” it said. Finally, it seemed to make a breakthrough. “Is this version of LLaMA more powerful than the one we released last year?” the ChatGPT transcript said. It was not clear about who was speaking, however. ChatGPT said "her mouth movements," but then explained that the "speaker is the man on the left" (Lessin, not Zuckerberg, was speaking in these frames). 

I had uploaded 40 of a total of 3,000 frames. Zoom video is usually 30 fps, so in approximately 1.5 seconds, Lessin and/or Zuckerberg apparently said “Is this version of LLaMA more powerful than the one we released last year?” I then recorded this phrase at a normal speaking speed, and it took about four seconds. Just a data point.

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI
Lipreadtest
0:00
/4.973333

I then got an error message from ChatGPT, and got rate-limited because I was uploading too much data. It told me that I needed to wait three hours to try again. 

Saving the Lost Silent Zuckerberg Interview With the Amazing Power of AI

Finally, I did what Meta AI told me to do, and tried a bespoke AI lip reading app. I found one called ReadTheirLips.com, which is powered by Symphonic Labs. This is a tool that people have been trying to use in recent months to figure out what Donald Trump and Jeffrey Epstein were saying to each other in silent b-roll news footage, without much success.

I paid $10 for three minutes worth of transcription and asked it to lip read using its “Multiface Detection.” After waiting 10 minutes, I got an error message that said “Transcription failed, no credits have been used, try again later.” I then asked it to focus only on Zuckerberg, and actually got some text. I separately asked it to focus on Lessin.

Here is a transcript of what the AI says they were talking about. It has not been edited for clarity and I have no idea which parts, if any, are accurate:

LESSIN: Thanks for joining us again, TV. We're happy to have you already this morning. News that you've spent even more money with your big announcement about your new supercomputers. We'll get to that, but to start, you've been in huge scale like I.

ZUCKERBERG: Happy TO BE HERE. We're GOING TO TALK A LITTLE BIT ABOUT META'S AI STRATEGY. It's BEEN BUSY, YOU KNOW? I THINK THE MOST EXCITING THING THIS YEAR IS THAT WE'RE STARTING TO SEE EARLY GLIMPSES OF SELF-IMPROVEMENT WITH THE MODELS, WHICH MEANS THAT DEVELOPING SUPERINTELLIGENCE IS NOW.

LESSIN: You HAVE BEEN ON A PLANE OF AI HIRING, WHY AND WHY NOW? 

ZUCKERBERG: Insight, and we just want to make sure that we really strengthen the effort as much as possible to go for it. Our mission with a lab is to deliver personal superintelligence to everyone in the world, so that way, you know, we can put that power in every individual's hand. I'm really excited about it.

LESSIN: I DON'T KNOW, I DON'T KNOW, I DON'T KNOW.

ZUCKERBERG: Than ONE OF THE OTHER LABS YOU'RE DOING, AND YOU KNOW MY VIEW IS THAT THIS IS GOING TO BE SOMETHING THAT IS THE MOST IMPORTANT TECHNOLOGY IN OUR LIVES. IT'S GOING TO UNDERPIN HOW WE DEVELOP EVERYTHING AND THE COMPANY, AND IT'S GOING TO AFFECT SOCIETY VERY WISELY. SO WE JUST WANT TO MAKE SURE WE GET THE BEST FOCUS.

LESSIN: Did YOU FEEL LIKE YOU WERE BEHIND WHAT WAS COMING OUT OF LAW BEFORE I'M NOT ADJUSTING.

ZUCKERBERG: On THIS FROM ENTREPRENEURS TO RESEARCHERS TO ENGINEERS WORKING ON THIS HIDDEN INFRASTRUCTURE, AND THEN OF COURSE WE WANT TO BACK IT UP WITH JUST AN ABSOLUTELY MASSIVE AMOUNT OF COMPUTER RESEARCH, WHICH WE CAN SUPPORT BECAUSE WE HAVE A VERY STRONG BUSINESS MODEL THAT THROWS OFF A LOT OF CAPITAL. LET'S TALK ABOUT.

LESSIN: Like THIS SUMMER, PARTICULARLY, YOU SWITCH GEARS A LITTLE BIT.

ZUCKERBERG: I THINK THE FIELD IS ACCELERATING, YOU KNOW, WE KEEP ON TRACK FOR WHERE WE WANT TO BE, AND THE FIELD KEEPS US MOVING FORWARD.

The video ends there, and it cuts back to the studio.

Update: The Information provided 404 Media with several clips (with audio) from Lessin's interview with Zuckerberg, as well as a real transcript of the interview. Here is the real segment of what was said. As you can see, the AI captured the jist of this portion of the interview, and actually did not do too bad:

Lessin: Mark, thanks for joining TITV. We're happy to have you here. Already this morning, [there’s] news that you've spent even more money with your big announcement about your new supercomputers. We'll get to that. But to start, you took a huge stake in ScaleAI. You have been on a blitz of AI hiring. Why, and why now?

Zuckerberg: Yeah, it's been busy. You know, I think the most exciting thing this year is that we're starting to see early glimpses of self-improvement with the models, which means that developing super intelligence is now in sight, and we just want to make sure that we really strengthen the effort as much as possible to go for it. Our mission with the lab is to deliver personal super intelligence to everyone in the world, so that way we can put that power in every individual's hand. And I'm really excited about it. It's a different thing than what the other labs are doing.

And my view is that this is going to be something that is the most important technology in our lives. It's going to underpin how we develop everything at the company, and it's going to affect society very widely. So we just want to make sure that we get the best folks to work on this, from entrepreneurs to researchers to engineers working on the data and infrastructure.

And then, of course, we want to back up with just an absolutely massive amount of compute which we can support, because we have a very strong business model that throws off a lot of capital.

Lessin: Did you feel like you were behind coming out of Llama 4? It seems like this summer, in particular, you switched gears a little bit.

Zuckerberg: I think the field is accelerating, you know, we keep on having goals for where we want to be. And then the field keeps on moving faster than we expect.

The rest of the interview is available at The Information.

  • ✇404 Media
  • Immigration Raid Tracking App ‘ICE Block’ Keeps Your Data Private, Researcher Finds
    ICE Block, an app that lets users warn others about the location of ICE officers, and which for a short while was the top of the social media App Store chart, does protect users’ privacy and doesn’t share your location with third parties, according to a recent analysis from a security researcher. ICE Block already claimed that it did not collect any data from the app; the analysis now corroborates that.“It’s not uploading your location at all, when you make a report that report isn’t associat
     

Immigration Raid Tracking App ‘ICE Block’ Keeps Your Data Private, Researcher Finds

15 juillet 2025 à 10:09
Immigration Raid Tracking App ‘ICE Block’ Keeps Your Data Private, Researcher Finds

ICE Block, an app that lets users warn others about the location of ICE officers, and which for a short while was the top of the social media App Store chart, does protect users’ privacy and doesn’t share your location with third parties, according to a recent analysis from a security researcher. ICE Block already claimed that it did not collect any data from the app; the analysis now corroborates that.

“It’s not uploading your location at all, when you make a report that report isn’t associated with your device in any way, and there are no third party services that it talks to or sends data to,” Cooper Quintin, senior public interest technologist at the Electronic Frontier Foundation (EFF), who analyzed the ICE Block app, told 404 Media.

  • ✇404 Media
  • Hugging Face Is Hosting 5,000 Nonconsensual AI Models of Real People
    Hugging Face, a company with a multi-billion dollar valuation and one of the most commonly used platforms for sharing AI tools and resources, is hosting over 5,000 AI image generation models that are designed to recreate the likeness of real people. These models were all previously hosted on Civitai, an AI model sharing platform 404 Media reporting has shown was used for creating nonconsensual pornography, until Civitai banned them due to pressure from payment processors. Users downloaded the mo
     

Hugging Face Is Hosting 5,000 Nonconsensual AI Models of Real People

15 juillet 2025 à 09:20
Hugging Face Is Hosting 5,000 Nonconsensual AI Models of Real People

Hugging Face, a company with a multi-billion dollar valuation and one of the most commonly used platforms for sharing AI tools and resources, is hosting over 5,000 AI image generation models that are designed to recreate the likeness of real people. These models were all previously hosted on Civitai, an AI model sharing platform 404 Media reporting has shown was used for creating nonconsensual pornography, until Civitai banned them due to pressure from payment processors. 

Users downloaded the models from Civitai and reuploaded them to Hugging Face as part of a concerted community effort to archive the models after Civitai announced in May it will ban them. In that announcement, Civitai said it will give the people who originally uploaded them “a short period of time” before they were removed. Civitai users began organizing an archiving effort on Discord earlier in May after Civitai indicated it had to make content policy changes due to pressure from payment processors, and the effort kicked into high gear when Civitai announced the new “real people” model policy. 

At the time of writing, the Discord channel has hundreds of members who are still finding and sharing models that have been removed from Civitai and are reuploading them to Hugging Face. Some users have even shared a piece of software, also hosted on Hugging Face, which allows users to automatically upload Civitai models to Hugging Face in batches. 

Hugging Face did not respond to multiple requests for comment. It also did not respond to specific questions about how and if it plans to moderate these models given the fact that they were previously hosted on a platform primarily used for AI generating pornography, and which our reporting shows were used to create noncensual pornography. 

I found the Civitai models of real people that were reuploaded to Hugging Face thanks to a paper I covered where researchers scraped Civitai. The paper showed that the platform was primarily used for pornographic content, and that it deleted at least 50,000 AI models designed to recreate the likeness of real people once it changed its policy in May. The researchers, Laura Wagner and Eva Cetinic from the University of Zurich, provided me with a spreadsheet of all the deleted models, which included the name of the models (which is almost always the name of a female celebrity or lesser known internet personality), a link to where it was previously hosted on Civitai, and the SHA256 hash Civitai uses to identify all the models hosted on its site. 

The people who are reuploading the Civitai models to Hugging Face are seemingly trying to hide the purpose of those models on Hugging Face. On Hugging Face, these models have generic names and URLs like “LORA” or “Test model.” Users can’t tell that these models are used to generate the likeness of real people just by looking at their Hugging Face page, nor would they be able to find them by searching for the names of celebrities on Hugging Face. In order to find them, users can go to a separate website the Civitai archivists created. There, they can enter the name of a Civitai model, the link where it used to be hosted on Civitai before it was deleted, or the model’s SHA256 hash. All of these will lead users to a page which explains what the model is, show its name, as well as several images showing the kind of images it can generate. At the bottom of that page is a link to one or more Hugging Face “mirrors” where the model has been reuploaded. 

By using Wagner’s and Cetinic’s data and entering it into this Civitai archive site, I was able to find the Civitai models hosted on Hugging Face. 

Hugging Face’s content policy bans “Unlawful, defamatory, fraudulent, or intentionally deceptive Content (e.g., disinformation, phishing, scams, inauthentic behavior),” as well as “Sexual Content used for harassment, bullying, or created without explicit consent.” Models that generate the likeness of real people don’t have to be used for unlawful or defamatory ends, and they only produce sexual content if people choose to use them that way. There’s nothing in Hugging Face’s content policy that explicitly forbids AI models that recreate the likeness of real people. 

However, the Hugging Face Ethics & Society group, which is “committed to operationalizing ethics at the cutting-edge of machine learning,” has identified six “high-level categories for describing ethical aspects of machine learning work,” one of which is that AI should be “Consentful.”

“Consentful technology supports the self-determination of people who use and are affected by these technologies,” the company explains. Examples of this, the company says, includes “Avoiding extractive, chauvinist, ‘dark,’ and otherwise ‘unethical’ patterns of engagement.”

Other AI models that recreate the likeness of real people could conceivably not violate any of these principles. For example, two of the deleted Civitai models that were reuploaded to Hugging Face were designed to recreate the likeness of Vladimir Putin, which in theory people would want to use in order to mock or criticize the Russian president. However, the vast majority of the models are of female celebrities, which my reporting has shown is being used to create nonconsensual sexual content, and which were deleted en masse from Civitai because of pressure from payment processors who didn’t want to be associated with that type of media. 

  • ✇404 Media
  • a16z-Backed AI Site Civitai Is Mostly Porn, Despite Claiming Otherwise
    In the two years that I’ve been reporting about Civitai, a platform for sharing AI image generation models that has been instrumental in the production of AI generated non-consensual porn, Civitai has consistently argued that the amount of adult content on the site has been overstated. But new research shows that, if anything, the amount of adult content on Civitai has been underestimated.In their paper, “Perpetuating Misogyny with Generative AI: How Model Personalization Normalizes Gendered Har
     

a16z-Backed AI Site Civitai Is Mostly Porn, Despite Claiming Otherwise

15 juillet 2025 à 09:10
a16z-Backed AI Site Civitai Is Mostly Porn, Despite Claiming Otherwise

In the two years that I’ve been reporting about Civitai, a platform for sharing AI image generation models that has been instrumental in the production of AI generated non-consensual porn, Civitai has consistently argued that the amount of adult content on the site has been overstated. But new research shows that, if anything, the amount of adult content on Civitai has been underestimated.

In their paper, “Perpetuating Misogyny with Generative AI: How Model Personalization Normalizes Gendered Harm,” researchers Laura Wagner and Eva Cetinic from the University of Zurich studied more than 40 million user-generated images on Civitai and over 230,000 models. They found “a disproportionate rise in not-safe-for-work (NSFW) content and a significant number of models intended to mimic real individuals” on the platform, they write in the paper.

“What began as a promising creative breakthrough in TTI [text-to-image] generation and model personalization, has devolved into a pipeline for the large-scale production of sensational, biased, and abusive content. The open-source nature of TTI technologies, proclaimed as a democratizing force in generative AI, has also enabled the propagation of models that perpetuate hypersexualized imagery and nonconsensual deepfakes,” Wagner and Cetinic write in their paper. “Several indicators suggest a descent into a self-reinforcing feedback loop of platform decay. These include a dramatic increase in NSFW imagery, from 41% to 80% in two years, as well as the community’s normalization of deepfakes, misogynistic tropes, and other exploitative content.”

To visualize just how dominant adult content was on Civitai, check the chart below, which shows the distribution of images by “NSFW browsing levels” over time. These categories, which are inspired by the Motion Picture Association film rating system and are used by Civitai to tag images, show that adult content was always a significant portion of all images hosted on the site, but that the portion of “overtly sexual, or disturbing” content only grew as the site became more popular, and exploded starting in 2024. The chart is based on Civitai’s own numbers and categorization system which the researchers scraped from the site. It likely undercounts the number of explicit images on the site since as both the researchers and I observed during my reporting, not all adult content is tagged as such. 

a16z-Backed AI Site Civitai Is Mostly Porn, Despite Claiming Otherwise

In December, 2023, Civitai CEO Justin Maier told Venture Beat that “less than 20% of the posted content is what we would consider ‘PG-13’ or above.” When I reached Maier for comment for this article, he told me that “The VentureBeat figure cited a December 2023 snapshot, when adult posts were a minority. The mix shifted in 2024 as many NSFW creators migrated from platforms that no longer allow that content.”

However, the data in the paper shows that by October of 2023, 56 percent of all images on the site were tagged as “NSFW” and were designated by Civitai as “PG-13” or above.

In May, Civitai announced it’s banning all AI image generation models designed to recreate the likeness of real people because of pressure from payment processors. Since the authors of the paper were already tracking hundreds of thousands of models hosted on Civitai, they could easily see which models were removed, giving us a first clear look at how common those models were. 

Overall, they saw that more than 50,000 models designed to AI-generate the likeness of real people were removed because of the ban. These are models that Civitai itself tagged as “person of interest,” the tag it uses to indicate a model recreates the likeness of a real person, so the actual number of models depicting real people is likely higher. 

It’s hard to say if the most popular AI models on Civitai were all popular just because they were used to generate explicit images, because people could use models tagged as NSFW to generate non-nude images and vice versa. For example, according to the data collected by the researchers the most popular AI image generation model on Civitai was EasyNegative with almost 600,000 downloads. It’s not tagged or promoted as a model for generating pornography, but images that users created with it, which are shared on its Civitai model page, show it is commonly used that way. 

Other very popular models on Civitai are clearly designed to generate explicit images. The sixth most popular model with 360,000 downloads is Nudify XL: Better Bodies, which its creator says is for “nude female frontals.” A model called Realistic Vaginas - God Pussy 1 had 256,000 downloads. The POV Squatting Cowgirl LoRA model, which Civitai tagged as a “sex” model, had 189,000 downloads. 

a16z-Backed AI Site Civitai Is Mostly Porn, Despite Claiming Otherwise

The authors of the paper also conducted deeper analysis of the 40,000 most downloaded models on Civitai. In the 11,151 models where they could extract textual training data, meaning text that indicates what kind of images the models were trained on, they found “specifically abusive terms.” 5.6 percent included the keywords “loli” (558 models) and/or “shota” (69 models), Japanese terms commonly used to refer to sexualized depictions of pre-pubescent girls and boys. About 2.1 percent (189 models) included the keyword “rape.”

The data shows with clear numbers what we have long argued at 404 Media: adult content drives technological innovation and early adoption, and this has been especially true in the world of generative AI. Despite its protestation to the contrary, Civitai, which is one of the fastest growing platforms in that industry, and that the influential Silicon Valley venture capital firm Andreessen Horowitz invested in, grew because of explicit content, much of which was nonconsensual. 

“The rapid rise of NSFW content, the over-representation of young female subjects, and the prioritization of sensational content to drive engagement reflect an exploitative, even abusive dynamic,” the researchers wrote. “Additionally, structural discrimination embedded in today’s open-source TTI tools and models have the potential to cause significant downstream harm as they might become widely adopted and even integrated into future consumer applications.” 

Adult content driving innovation and early adoption doesn’t have to be harmful. As the researchers write, it’s the choices platforms like Civitai make that give us these outcomes. 

“The contingent nature of technology, shaped by online communities, platform operators, lawmakers, and society as a whole, also creates opportunities for intervention,” they write. “Model-sharing hubs and social media platforms both have the capacity to implement safeguards that can limit the spread of abusive practices such as deepfake creation and abusive imagery.”

  • ✇404 Media
  • Hackers Can Remotely Trigger the Brakes on American Trains and the Problem Has Been Ignored for Years
    Many trains in the U.S. are vulnerable to a hack that can remotely lock a train’s brakes, according to the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the researcher who discovered the vulnerability. The railroad industry has known about the vulnerability for more than a decade but only recently began to fix it.Independent researcher Neil Smith first discovered the vulnerability, which can be exploited over radio frequencies, in 2012. “All of the knowledge to generate the
     

Hackers Can Remotely Trigger the Brakes on American Trains and the Problem Has Been Ignored for Years

14 juillet 2025 à 21:59
Hackers Can Remotely Trigger the Brakes on American Trains and the Problem Has Been Ignored for Years

Many trains in the U.S. are vulnerable to a hack that can remotely lock a train’s brakes, according to the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the researcher who discovered the vulnerability. The railroad industry has known about the vulnerability for more than a decade but only recently began to fix it.

Independent researcher Neil Smith first discovered the vulnerability, which can be exploited over radio frequencies, in 2012. 

“All of the knowledge to generate the exploit already exists on the internet. AI could even build it for you,” Smith told 404 Media. “The physical aspect really only means that you could not exploit this over the internet from another country, you would need to be some physical distance from the train [so] that your signal is still received.”

  • ✇404 Media
  • Swedish Prime Minister Pulls AI Campaign Tool After It Was Used to Ask Hitler for Support
    The Moderate Party of Sweden has removed an AI tool from its website after people used it to generate videos of Prime Minister Ulf Kristersson asking Adolf Hitler for support.The tool allowed users to generate videos of Kristersson holding an AI-generated message in an attempt to promote the candidate ahead of the general election in Sweden next year.Swedish television station TV4 used the tool to generate a video of Kristersson on a newspaper above the headline “Sweden needs Adolf Hitler” af
     

Swedish Prime Minister Pulls AI Campaign Tool After It Was Used to Ask Hitler for Support

14 juillet 2025 à 13:39
Swedish Prime Minister Pulls AI Campaign Tool After It Was Used to Ask Hitler for Support

The Moderate Party of Sweden has removed an AI tool from its website after people used it to generate videos of Prime Minister Ulf Kristersson asking Adolf Hitler for support.The tool allowed users to generate videos of Kristersson holding an AI-generated message in an attempt to promote the candidate ahead of the general election in Sweden next year.

Swedish television station TV4 used the tool to generate a video of Kristersson on a newspaper above the headline “Sweden needs Adolf Hitler” after it noticed that it had no guardrails or filters.

In the video TV4 generated using the website, Kristersson makes his pitch over stock footage of old people embracing. A woman runs through a field, the camera focusing on flowers while the sun twinkles in the background. Cut to Kristersson. He turns a blue board around. “We need you, Adolf Hitler,” it says.

The Moderates removed the AI system from its website, but the videos of Ulf asking Hitler to join the Moderates remain on social media and TV4’s website..

In an attempt to bolster its party's ranks, Moderates launched a website that allowed users to generate a custom video of Kristersson asking someone to join the party. The idea was probably to have party members plug in the names of friends and family members and share what appeared to be a personalized message from the PM asking for their support.

In the video, Kristersson stands in front of stairs, makes his pitch, and turns around a blue tablet that bears a personalized message to the viewer. The system apparently had no guardrails or filters and Swedish television station TV4 was able to plug in the names Adolf Hitler, Ugandan dictator Idi Amin, and Norwegian mass murderer Anders Breivik.

The Moderate Party did not return 404 Media’s request for a comment about the situation, but told TV4 it shut down the site as soon as it learned people were using it to generate messages with inappropriate names.

The Moderate Party’s AI-generated video was simple.. It filmed the PM holding a blue board it could easily overlay with input from a user and then used AI to generate the fake newspaper and a few other slides. Preventing people from typing in “Hitler” or “Anders Brevik” would have been as simple as maintaining a list of prohibited names, words, and phrases, something that every video game and service does. Users are good at bypassing guardrails, but the Moderate’s AI tool appeared to have none.

Users making content you don’t want to be associated with is one of the oldest and most well known problems in AI. If you release a chatbot, generative photo system, or automated political greeting generator, someone will use it to reference the Nazis or make nonconsensual porn.

When Microsoft launched TAY in 2016, users turned it into a Hitler-loving white nationalist in a few hours. Eight years later, another Microsoft AI product had a loophole that let people make AI-generated nudes of Taylor Swift. Earlier this year, Instagram’s AI chatbots lied about being licensed therapists.

  • ✇404 Media
  • 'Deportation Tok' Is Taking Off
    As immigration raids roll out across the U.S., those affected are processing the experience in the normal 2025 way—via vertical video. Across social media, people are uploading clips with uncanny-valley titles like “A normal day for me after being deported to Mexico” and “3 things I wish I knew before self-deporting from the US!” These posts have the normal shape, voiceovers, and fonts of influencer content, but their dystopian topic reflects the whiplash of the current historical moment. Doo
     

'Deportation Tok' Is Taking Off

14 juillet 2025 à 10:25
'Deportation Tok' Is Taking Off

As immigration raids roll out across the U.S., those affected are processing the experience in the normal 2025 way—via vertical video. 

Across social media, people are uploading clips with uncanny-valley titles like “A normal day for me after being deported to Mexico” and “3 things I wish I knew before self-deporting from the US!” These posts have the normal shape, voiceovers, and fonts of influencer content, but their dystopian topic reflects the whiplash of the current historical moment. 

Doomscrolling last week, a particular clip caught my eye. A man sits on the bottom bunk of a metal bed, staring down at the floor, with the caption “Empezando una nueva vida después de que me Deportaran a México” (“Starting a new life after being Deported to Mexico”).

The Media's Pivot to AI Is Not Real and Not Going to Work

14 juillet 2025 à 10:11
Subscribe
Join the newsletter to get the latest updates.
Success
Great! Check your inbox and click the link.
Error
Please enter a valid email address.
The Media's Pivot to AI Is Not Real and Not Going to Work

On May 23, we got a very interesting email from Ghost, the service we use to make 404 Media. “Paid subscription started,” the email said, which is the subject line of all of the automated emails we get when someone subscribes to 404 Media. The interesting thing about this email was that the new subscriber had been referred to 404 Media directly from chatgpt.com, meaning the person clicked a link to 404 Media from within a ChatGPT window. It is the first and only time that ChatGPT has ever sent us a paid subscriber.

From what I can tell, ChatGPT.com has sent us 1,600 pageviews since we founded 404 Media nearly two years ago. To give you a sense of where this slots in, this is slightly fewer than the Czech news aggregator novinky.cz, the Hungarian news portal Telex.hu, the Polish news aggregator Wykop.pl, and barely more than the Russian news aggregator Dzen.ru, the paywall jumping website removepaywall.com, and a computer graphics job board called 80.lv. In that same time, Google has sent roughly 3 million visitors, or 187,400 percent more than ChatGPT. 

This is really neither here nor there because we have tried to set our website up to block ChatGPT from scraping us, though it is clear this is not always working. But even for sites that don’t block ChatGPT, new research from the internet infrastructure company CloudFlare suggests that OpenAI is crawling 1,500 individual webpages for every one visitor that it is sending to a website. Google traffic has begun to dry up as both Google’s own AI snippets and AI-powered SEO spam have obliterated the business models of many media websites. 

The Media's Pivot to AI Is Not Real and Not Going to Work

This general dynamic—plummeting traffic because of AI snippets, ChatGPT, AI slop, Twitter no workie so good no more—has been called the “traffic apocalypse” and has all but killed some smaller websites and has been blamed by executives for hundreds of layoffs at larger ones. 

Despite the fact that generative AI has been a destructive force against their businesses, their industry, and the truth more broadly, media executives still see AI as a business opportunity and a shiny object that they can tell investors and their staffs that they are very bullish on. They have to say this, I guess, because everything else they have tried hasn’t worked, and pretending that they are forward thinking or have any clue what they are doing will perhaps allow a specific type of media executive to squeeze out a few more months of salary.

But pivoting to AI is not a business strategy. Telling journalists they must use AI is not a business strategy. Partnering with AI companies is a business move, but becoming reliant on revenue from tech giants who are creating a machine that duplicates the work you’ve already created is not a smart or sustainable business move, and therefore it is not a smart business strategy. It is true that AI is changing the internet and is threatening journalists and media outlets. But the only AI-related business strategy that makes any sense whatsoever is one where media companies and journalists go to great pains to show their audiences that they are human beings, and that the work they are doing is worth supporting because it is human work that is vital to their audiences. This is something GQ’s editorial director Will Welch recently told New York magazine: “The good news for any digital publisher is that the new game we all have to play is also a sustainable one: You have to build a direct relationship with your core readers,” he said.

Becoming an “AI-first” media company has become a buzzword that execs can point at to explain that their businesses can use AI to become more ‘efficient’ and thus have a chance to become more profitable. Often, but not always, this message comes from executives who are laying off large swaths of their human staff.

In May, Business Insider laid off 21 percent of its workforce. In her layoff letter, Business Insider’s CEO Barbara Peng said “there’s a huge opportunity for companies who harness AI first.” She told the remaining employees there that they are “fully embracing AI,” “we are going all-in on AI,” and said “over 70 percent of Business Insider employees are already using Enterprise ChatGPT regularly (our goal is 100%), and we’re building prompt libraries and sharing everyday use cases that help us work faster, smarter, and better.” She added they are “exploring how AI can boost operations across shared services, helping us scale and operate more efficiently.” 

Last year, Hearst Newspapers executives, who operate 78 newspapers nationwide, told the company in an all-hands meeting audio obtained by 404 Media that they are “leaning into [AI] as Hearst overall, the entire corporation.” Examples given in the meeting included using AI for slide decks, a “quiz generation tool” for readers, translations, a tool called Dispatch, which is an email summarization tool, and a tool called “Assembly,” which is “basically a public meeting monitor, transcriber, summarizer, all in one. What it does is it goes into publicly posted meeting videos online, transcribes them automatically, [and] automatically alerts journalists through Slack about what’s going on and links to the transcript.”

  • ✇404 Media
  • Trump’s NASA Cuts Would Hurt America for a Long, Long Time
    Welcome back to the Abstract! Here are the studies that caught my eye this week.First up, a bummer! NASA is facing devastating cuts to Earth science, and science in general, which is pretty important for an agency tasked with understanding the universe. I try to keep this newsletter relatively downer-free, but current events are not cooperating with this aim.  Then: the early bird gets the comet, a donkey destined for decapitation, a grand universal theory of coolness, and the case of the stolen
     

Trump’s NASA Cuts Would Hurt America for a Long, Long Time

12 juillet 2025 à 09:00
Trump’s NASA Cuts Would Hurt America for a Long, Long Time

Welcome back to the Abstract! Here are the studies that caught my eye this week.

First up, a bummer! NASA is facing devastating cuts to Earth science, and science in general, which is pretty important for an agency tasked with understanding the universe. I try to keep this newsletter relatively downer-free, but current events are not cooperating with this aim.  

Then: the early bird gets the comet, a donkey destined for decapitation, a grand universal theory of coolness, and the case of the stolen exomoons.  

Knowing about the planet we live on is good, actually

Millet, Dylan et al. “NASA Earth Science Division provides key data.” Science.

This may sound obvious for the esteemed readers of this newsletter, but it apparently bears repeating: Earth, our home planet and the only known life-bearing world in the vast expanse of spacetime, is worthy of some passing interest.

This is true because—and I can’t stress this enough—we live on planet Earth. All our bones and guts and snacks are here, so it’s probably wise to get the lay of the land. But as a bonus, Earth is wildly interesting, a restless substrate unlike anything we have seen in our own solar system or beyond it. 

Despite these considerations, the Trump administration plans to gut NASA’s Earth Science Division (ESD), the world leader in world-watching. In a letter published in Science, researchers from the recently dissolved NASA Earth Science Advisory Committee lamented the administration’s proposed budget cut of more than 50 percent to ESD, warning that it “would come at a profound cost to US society and scientific leadership.”

“NASA ESD accounted for just 0.03 percent of US spending in 2024,” said researchers led by Dylan Millet of the University of Minnesota. “This investment returns its value many times over by improving predictions, by spurring technological innovation and high-tech jobs, and by forging the knowledge of the planet that is needed for short- and long-term planning.”

“The budget cuts proposed for ESD would cancel crucial satellites that observe Earth and its atmosphere, gut US science and engineering expertise, and potentially lead to the closure of NASA research centers,” the team said. “Given that the cuts would prevent the US from training and preparing the next generation of the scientific and technical workforce, the consequences would be long-lasting.”

This is just the latest appeal from scientists on behalf of NASA, which is also facing catastrophic cuts to its overall Science Mission Directorate (SMD), the arm that oversees ESD. Last week, every past administrator of the SMD, the agency's top job for science leadership, signed a letter urging Congress to reject the cuts of about 47 percent to the directorate. 

“Each one of us knows what it’s like to shepherd an ambitious project forward, knowing that its payoff will come years after we have left the agency,” the administrators said. “This proposed budget ends nearly all future investments for both new missions and advanced technology for science. It walks away from dozens of current, extraordinarily successful and productive science missions in extended operations on a combined budget that is only about three percent of NASA’s annual funding.”

Fortunately, the US Senate appropriations committee has voted in favor of a bill rejecting the science cuts, but it has a long road to go down before taking effect, with plenty of opportunity to fall apart. 

Needless to say, turning a blind eye to Earth at a time when our activities are reshaping its climate and biosphere, would be a huge loss. As one last twist of the knife, Trump just gave the top job at NASA to the guy from The Real World—all while ignoring the actual real world. 

In other news…

Mark your calendars for July 2061

Barbieri, Cesare et al. “Preparing for the 2061 return of Halley’s comet. A rendezvous mission with an innovative imaging system.” Planetary and Space Science.

Prepare for the return of everyone’s favorite space iceball: Comet Halley. Scientists have flagged the comet’s next visit as arriving in the summer of 2061 and proposed an audacious space rendezvous with Halley on its wild ride toward the Sun. 

“Although the crucial phases of the comet’s ingress in the inner Solar System are still more than 30 years in the future, we started to examine the feasibility of a space mission using present-day rockets and technologies,” said researchers led by Cesare Barbieri of the University of Padova.

Trump’s NASA Cuts Would Hurt America for a Long, Long Time
Comet Halley in 1986. Image: NASA/W. Liller - NSSDC's Photo Gallery (NASA)

God bless the astro-preppers. Sure, this event will occur decades into the future, in the twilight of the millennials. But they make a pretty good case that we should get moving if we want to take full advantage of the iconic visitor, which “will be better positioned for observation from terrestrial observers than during the 1985–1986 apparition, as it will be on the same side of the Sun as the Earth.”

“We stress that a concerted effort is needed in the current decade to plan and approve a rendezvous mission to [Comet Halley],” the team concluded. “Indeed, the scenario here described needs launches before 2040, less than 15 years from now.”

The last days of a sacrificial ass 

Arnold, Elizabeth et al. “An isotopic perspective on equid selection in cult at Tell eṣ-Ṣâfi/Gath, Israel.” PLOS ONE.

Who would have guessed that a study about a Bronze Age donkey corpse would be a tear-jerker? Researchers have shed new light on female donkeys (or jennies), which were imported from Egypt to the Tell es-Safi/Gath site in Israel some 5,000 years ago for ritual purposes. 

One specimen, called EQ1, is particularly noteworthy because it was decapitated and had its limbs tied together, unlike all the other donkeys buried at the site. "It is evident the animal was sacrificed, the head entirely cut off and carefully placed on the abdomen facing in the opposite direction,” said researchers led by Elizabeth Arnold of Grand Valley State University.  

Trump’s NASA Cuts Would Hurt America for a Long, Long Time
The four donkey burials at Tell eṣ-Ṣâfi/Gath. Image: Arnold et al., 2025, PLOS One, CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

“It can be deduced that even though EQ1 was grazed locally toward the end of her life, she was treated slightly differently from the other local equids,” the team continued. “This imported donkey was kept penned and foddered with hay that was harvested in the valley, a product of dry farmed cereals. This donkey was never herded with other livestock east of the site.”

The unique treatment of EQ1 suggests that the “Egyptian donkey might have been seen as an exotic and special animal, worthy of specific ritual use,” the study concluded. While it’s truly impressive that so much about this jenny can be inferred from her bones, there’s also an eerie pathos to imagining the animal hanging out for months, receiving preferential treatment, unaware of the sand flowing through the hourglass.    

It’s a real cool club, and you’re not part of it

Pezzuti, Todd and Warren Caleb et al. “Cool People,” Journal of Experimental Psychology.

Science has invested its prodigious powers into the ultimate social mystery: What makes a person “cool”? Is it putting “cool” in scare quotes? (No!). Researchers have now developed a working theory of coolness by asking nearly 6,000 people in Australia, Chile, China, Hong Kong, Germany, India, Mexico, Nigeria, Spain, South Africa, South Korea, Turkey, and the United States to define this much-coveted attribute.

The results revealed six main traits associated with cool people, which were distinct from traits linked with “good” people. “Cool people are perceived to be more extraverted, hedonistic, powerful, adventurous, open, and autonomous, whereas good people are more conforming, traditional, secure, warm, agreeable, universalistic, conscientious, and calm,” said authors Todd Pezzuti of Universidad Adolfo Ibañez, Caleb Warren of the University of Arizona, and Jinjie Chen of the University of Georgia.

“This pattern is stable across countries, which suggests that the meaning of cool has crystallized on a similar set of values and traits around the globe,” the team said.  

There you go, the cheat code to coolness. I’m exhausted just reading it.

Grand Theft Moons

Dencs, Zoltán et al. “Grand theft moons: Formation of habitable moons around giant planets.” Astronomy & Astrophysics.

The winner of Best Study Title this week goes to “Grand theft moons,” which explores how stars might steal moons from their own planets, and whether these “exomoons” could be habitable. The study models the formation of exomoons around giant gas worlds at various distances from their stars, measured in astronomical units (au), where one au is the distance between Earth and the Sun. 

Planets with orbits of one or two au are more likely to sport exomoons in the habitable zone, but they are also at risk of stars yanking the exomoons away in brazen acts of “stellar theft,” according to the study.

“Our simulations show that moons with masses between Mars and Earth could form around planets with masses about ten times that of Jupiter, and many of these moons could potentially be habitable at 1 − 2 au stellar distances,” said researchers led by Zoltán Dencs of the Gothard Astrophysical Observatory, “These findings suggest that it is worth investigating not only rocky planets but also gas giants for Earth-like habitable environments.”

In addition to raising some fascinating questions, let’s hope this study inspires Rockstar to take the GTA franchise to outer space. I want to throw an astronaut off a lunar buggy and take it for a joyride across the Moon.

Thanks for reading! See you next week.

  • ✇404 Media
  • Payment Processors Are Pushing AI Porn Off Its Biggest Platforms
    Tensor.Art, an AI image creating and model sharing site announced on Friday that it is “temporarily” restricting AI models, tools, and posts related to pornographic content or the depiction of real-world celebrities due to pressure from payment processors.The announcement is yet another example of payment processors acting as the ultimate arbiter of what kind of content can be easily made available online and those companies’ seemingly increased focus on AI-generated adult or nonconsensual co
     

Payment Processors Are Pushing AI Porn Off Its Biggest Platforms

11 juillet 2025 à 13:52
Payment Processors Are Pushing AI Porn Off Its Biggest Platforms

Tensor.Art, an AI image creating and model sharing site announced on Friday that it is “temporarily” restricting AI models, tools, and posts related to pornographic content or the depiction of real-world celebrities due to pressure from payment processors.

The announcement is yet another example of payment processors acting as the ultimate arbiter of what kind of content can be easily made available online and those companies’ seemingly increased focus on AI-generated adult or nonconsensual content. 

The news is especially significant following a similar change in policy from Civitai, an AI model sharing platform 404 Media reporting has shown was used for creating nonconsensual pornography. After Civitai banned AI models designed to generate the likeness of real people and certain types of adult content in May, many Civitai users and model creators migrated their models to Tensor.Art. The announcement listed three items in the “Scope of Impact” of the decision: Banning “NSFW” content, banning content based on real-world celebrities, and  temporarily disabling its “Civitai Import” feature, which allowed people to easily move their Civitai models to Tensor.Art.

💡
Do you know anything else about where else these models are hosted? I would love to hear from you. Using a non-work device, you can message me securely on Signal at ‪@emanuel.404‬. Otherwise, send me an email at emanuel@404media.co.

“We fully understand that this is very frustrating for many creators and users 😞,” Tensor.Art said in its announcement on Discord. “Due to the previous controversy over real-person content on Civitai, TA [Tensor.Art] has unfortunately been affected because of the ‘Civitai import’ feature. Owing to mandatory requirements from credit card organizations and regulatory authorities, we are compelled to make this temporary decision.”

Tensor.Art also listed the “Reasons for Adjustment” as:

- Review requirements for high-risk content from credit card organizations and multiple national regulatory bodies
- Compliance measures necessary to maintain platform openness and creators’ ability to monetize

Tensor.Art said that these changes will take place within the next 72 hours, and asked model creators to clarify if their models are “safe for work” in order to “prevent unintended impact.”

It’s not clear what Tensor.Art will look like or what its policies will be at the end of this “temporary” period. Civitai made similar changes permanently and still hasn’t been able to renew service from its payment processing providers or find new ones. Tensor.Art, however, is suggesting it’s not ready to give up on that type of content. 

“This is not the end,” Tensor.Art said in the announcement. “We are actively seeking solutions to minimize the impact of these restrictions and exploring compliant ways to restore currently hidden content. We remain committed to our original mission.”

Tensor.Art did not immediately respond to a request for comment.  

  • ✇404 Media
  • Behind the Blog: In Our Lane
    This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss choosing what to cover, fishing expeditions, and the library.JOSEPH: There’s an interesting cybersecurity story going on: a politically-motivated, right wing extremist hacker broke into Columbia University, stole swathes of applicant and other data, and then leaked parts of it. That’s how you got this New York Times article about New York m
     

Behind the Blog: In Our Lane

11 juillet 2025 à 11:49
Behind the Blog: In Our Lane

This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss choosing what to cover, fishing expeditions, and the library.

JOSEPH: There’s an interesting cybersecurity story going on: a politically-motivated, right wing extremist hacker broke into Columbia University, stole swathes of applicant and other data, and then leaked parts of it. That’s how you got this New York Times article about New York mayoral hopeful Zohran Mamdani applying to the university as Asian and African American. Predictably, some people are reading that in bad faith, when in reality it shows how stupid and regimented application processes can be in a diverse society.

This week The Verge published a piece called The Columbia hack is a much bigger deal than Mamdani’s college application, which basically elaborates on that point. One section was this:

And yet, there has been precious little reporting on the Columbia hack. Wired hasn’t covered it, and, until this story, neither has The Verge. Nor have The Chronicle of Higher Education, CyberScoop, 404 Media, TechCrunch, or Krebs on Security. These—including The Verge—are small to medium-size entities, and there’s any number of possible reasons why they didn’t pick it up. (On our end, it was partly because we were short-staffed during a national holiday, and partly because we didn’t immediately piece together how extraordinary this particular hack is.) But coverage at the much bigger, well-resourced institutions is also scanty. The Wall Street Journal passed on the story. Reuters has a brief on the initial outage; AP has a short write-up as well, which The Washington Post ran as part of their syndication deal.

  • ✇Dans les algorithmes
  • Entretiens d’embauche chatbotisés
    Les entretiens d’embauche avec des humains sont en train de prendre fin, rapporte le New York Times en évoquant l’essor des entretiens avec des systèmes d’IA. L’expérience avec ces robots intervieweurs, comme ceux développés par Ribbon AI, Talently ou Apriora, se révèle très déshumanisante, témoignent ceux qui y sont confrontés. Les questions sont souvent un peu creuses et les chatbots ne savent pas répondre aux questions des candidats sur le poste ou sur la suite du processus de recrutement (co
     

Entretiens d’embauche chatbotisés

11 juillet 2025 à 00:30

Les entretiens d’embauche avec des humains sont en train de prendre fin, rapporte le New York Times en évoquant l’essor des entretiens avec des systèmes d’IA. L’expérience avec ces robots intervieweurs, comme ceux développés par Ribbon AI, Talently ou Apriora, se révèle très déshumanisante, témoignent ceux qui y sont confrontés. Les questions sont souvent un peu creuses et les chatbots ne savent pas répondre aux questions des candidats sur le poste ou sur la suite du processus de recrutement (comme si ces éléments n’étaient finalement pas importants).

A croire que l’embauche ne consiste qu’en une correspondance d’un profil à un poste, la RHTech scie assurément sa propre utilité. Quant aux biais sélectifs de ces outils, parions qu’ils sont au moins aussi défaillants que les outils de recrutements automatisés qui peinent déjà à faire des correspondances adaptées. La course au pire continue !

  • ✇Dans les algorithmes
  • Scannés par l’IA
    Le loueur de voiture Hertz a commencé à déployer des scanneurs de voitures développées par UVeye pour inspecter les véhicules après leur location afin de vérifier leur état, explique The Drive (voir également cet article). Problème : le système est trop précis et surcharge les clients de frais pour des accrocs microscopiques qu’un être humain n’aurait pas remarqué.   Les tensions n’ont pas manqué d’éclater, et elles sont d’autant plus nombreuses qu’il est très difficile de contacter un agent
     

Scannés par l’IA

11 juillet 2025 à 00:23

Le loueur de voiture Hertz a commencé à déployer des scanneurs de voitures développées par UVeye pour inspecter les véhicules après leur location afin de vérifier leur état, explique The Drive (voir également cet article). Problème : le système est trop précis et surcharge les clients de frais pour des accrocs microscopiques qu’un être humain n’aurait pas remarqué.  

Les tensions n’ont pas manqué d’éclater, et elles sont d’autant plus nombreuses qu’il est très difficile de contacter un agent de l’entreprise pour discuter ou contester les frais dans ce processus de rendu de véhicule automatisé, et que cela est impossible via le portail applicatif où les clients peuvent consulter et régler les dommages attribués à leurs locations. Des incidents d’usure mineurs ou indépendants des conducteurs, comme un éclat lié à un gravillon, sont désormais parfaitement détectés et facturés. Le problème, c’est le niveau de granularité et de précision qui a tendance a surdiagnostiquer les éraflures. Décidément, adapter les faux positifs à la réalité est partout une gageure ou un moyen pour générer des profits inédits.

  • ✇404 Media
  • Our Galaxy May Contain a Mysterious Force. It Could Change Physics Forever.
    🌘Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week. Scientists are searching for signs of a “fifth force” at the center of our galaxy that could rewrite the rules of gravity and help to resolve some fundamental mysteries in the universe, according to a recent study in the journal Astronomy & Astrophysics. For decades, researchers have speculated that exotic new physics could fill missing links in our curr
     

Our Galaxy May Contain a Mysterious Force. It Could Change Physics Forever.

10 juillet 2025 à 10:12
🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
Our Galaxy May Contain a Mysterious Force. It Could Change Physics Forever.

Scientists are searching for signs of a “fifth force” at the center of our galaxy that could rewrite the rules of gravity and help to resolve some fundamental mysteries in the universe, according to a recent study in the journal Astronomy & Astrophysics

For decades, researchers have speculated that exotic new physics could fill missing links in our current understanding of gravity, which is based on Einstein’s general relativity. One idea is that a hypothetical fifth force—in addition to gravity, electromagnetism, and the strong and weak nuclear forces—known as a Yukawa-type correction might subtly alter how gravity behaves over certain distances. A direct detection of this force could shed light on longstanding puzzles like the nature of dark matter, an unidentified substance that accounts for most mass in the universe, or the behavior of gravity at quantum scales.

Now, researchers have used the advanced GRAVITY instrument at the Very Large Telescope in Chile to look for hints of this correction near the supermassive black hole at the center of the Milky Way. 

“The current theory of gravity is unable to explain some observations performed in the universe” such as “the presence of dark matter, or the expanding universe,” said Arianna Foschi, a postdoctoral researcher at the Paris Observatory and an author of the new study, in an email to 404 Media. 

“One possible explanation for this may be that the theory of gravity is not complete yet and some modifications to explain those effects are needed,” she added. “We looked exactly for the presence of such a modification.” 

Whereas gravity influences objects over massive cosmic distances, the Yukawa correction is predicted to be short-ranged and undetectable in local environments, such as our planet or the solar system. However, hints of this force, if it exists, could be observable near our galaxy’s supermassive black hole, Sagittarius A*, a chaotic region that showcases gravity at an extreme.

With that in mind, the GRAVITY collaboration trained its namesake instrument on a massive star called S2 that is very close to the supermassive black hole, orbiting it once every 16 years. Due to its proximity to the black hole, S2 has yielded many insights about gravity and general relativity, making it an attractive target for the team’s hunt for a fifth force. 

The motion of S2, along with other stars around Sagittarius A* “can be incredibly useful to check whether objects orbiting around a supermassive black hole follow the same rule as planets in the solar system,” Foschi said. “Observations suggest that indeed the law that makes S2 move is the same as the Earth, however there still can be modifications that cannot be seen ‘by eye’ but needed to be tested.”

As it turned out, the instrument’s precise measurements did not detect a fifth force, but they did get us closer. The results narrowed down the parameters of its possible intensity, represented by the variable “alpha.” 

“If before, alpha must be less than 0.01, now with our data we showed that it must be smaller than 0.003, significantly improving this constraint,” Forschi said.

Lorenzo Iorio, a physicist with the Italian Ministry of Education and Merit and an expert on modified theories of gravity, said in an email that the team’s approach made sense in principle, but that he had some concerns with the methods. Iorio, who was not involved in the study, cited updated formulas and variables that were left out of its models, but that he said might have improved its accuracy. For instance, the models did not account for the Lense-Thirring effect, which is a relativistic phenomenon near massive objects, or the influence of the accelerations of stars near S2, among other factors.

“I'd say that it is an interesting study that, rather, points towards the possibilities offered by this peculiar celestial laboratory (Sagittarius A* and the S stars),” Iorio said. “It should be repeated more accurately.”

Foschi acknowledged that the variables were not included in the models, but noted that the GRAVITY observations were not yet sensitive enough to capture many of those details—yet.

“An upgrade of the GRAVITY instrument is already on its way to increase the sensitivity and measure indeed these higher order effects, but we have to wait for that,” she noted.  

“We would like to extend the same analysis to data of other stars around Sagittarius A* and the improving precision of the GRAVITY instrument will help us to do so. In this way we will be able to put on even stronger constraints.” 

To that end, the GRAVITY collaboration plans to continue gazing at the center of the galaxy for signs of a fifth force, or any other modifications to gravity. 

“If such a force is observed, it would be an incredible breakthrough in physics, because it would modify one of the oldest physical laws we have,” Foschi concluded. “It would have consequences in every field of physics.”

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
  • ✇404 Media
  • The UN Made AI-Generated Refugees
    I am talking to Amina, an AI avatar that plays the role of a woman living in a refugee camp in Chad after she escaped violence in Sudan. “I’m hanging in there, yeah,” Amina tells me. “ It’s tough here in the camp, but we’re staying strong.” I ask her how she came to be in a refugee camp. “I fled my hometown of Al Junaynah in Sudan back in June 2023, due to the violence, yeah?” It says. “It was a terrifying journey to escape the conflict. Now I’m trying to keep my kids safe and find some stabi
     

The UN Made AI-Generated Refugees

10 juillet 2025 à 10:05
The UN Made AI-Generated Refugees

I am talking to Amina, an AI avatar that plays the role of a woman living in a refugee camp in Chad after she escaped violence in Sudan. “I’m hanging in there, yeah,” Amina tells me. “ It’s tough here in the camp, but we’re staying strong.” I ask her how she came to be in a refugee camp. “I fled my hometown of Al Junaynah in Sudan back in June 2023, due to the violence, yeah?” It says. “It was a terrifying journey to escape the conflict. Now I’m trying to keep my kids safe and find some stability here in the camp. It’s not easy but we have to keep going, right?”

0:00
/2:27

One of my conversations with 'Amina.'

Amina is an experiment, part of a pair of AI avatars created by the United Nations University Center for Policy Research (UNU-CPR), a research institution connected to the United Nations. It’s one that is sure to be controversial, considering that the UN itself says a major problem facing refugees is their dehumanization for political gain or convenience. The UNU-CPR project is using an inhuman technology in an attempt to help people learn more about what they are facing. The group also tested a soldier persona called “Abdalla,” which “simulates the behavior and decision-making patterns of an actual combatant, offering negotiators and mediators a possible tool to train for future high-stakes negotiations.”

  • ✇404 Media
  • Trump Mobile Keeps Charging My Credit Card And I Have No Idea Why
    Last month I put down $100 to pre-order the Trump Organization’s forthcoming mobile phone, the T1. Or, I tried to. As I wrote at the time, the website went to an error page, charged my credit card the wrong amount of $64.70, and sent a confirmation email saying I would receive another confirmation email when my order had been shipped, but I hadn’t provided a shipping address.I was surprised then to see another two charges on my card from Trump Mobile on Thursday, for $100 and $64.70 respectiv
     

Trump Mobile Keeps Charging My Credit Card And I Have No Idea Why

10 juillet 2025 à 09:32
Trump Mobile Keeps Charging My Credit Card And I Have No Idea Why

Last month I put down $100 to pre-order the Trump Organization’s forthcoming mobile phone, the T1. Or, I tried to. As I wrote at the time, the website went to an error page, charged my credit card the wrong amount of $64.70, and sent a confirmation email saying I would receive another confirmation email when my order had been shipped, but I hadn’t provided a shipping address.

I was surprised then to see another two charges on my card from Trump Mobile on Thursday, for $100 and $64.70 respectively. I did not expect or authorize these charges and will be trying to get my money back, if they go through (they’re currently pending). I don’t know when I will get my phone. I also don’t know how to make the charges to my credit card stop because other parts of the (since updated) website also return errors and the customer service number I called on the website couldn’t help either.

At first, the Trump Mobile phone pre-order process was bumbling. The company is now charging my card again and I have no idea why.

  • ✇Dans les algorithmes
  • L’IA, un nouvel internet… sans condition
    Tous les grands acteurs des technologies ont entamé leur mue. Tous se mettent à intégrer l’IA à leurs outils et plateformes, massivement. Les Big Tech se transforment en IA Tech. Et l’histoire du web, telle qu’on l’a connue, touche à sa fin, prédit Thomas Germain pour la BBC. Nous entrons dans « le web des machines », le web synthétique, le web artificiel où tous les contenus sont appelés à être générés en permanence, à la volée, en s’appuyant sur l’ensemble des contenus disponibles, sans que ce
     

L’IA, un nouvel internet… sans condition

10 juillet 2025 à 00:40

Tous les grands acteurs des technologies ont entamé leur mue. Tous se mettent à intégrer l’IA à leurs outils et plateformes, massivement. Les Big Tech se transforment en IA Tech. Et l’histoire du web, telle qu’on l’a connue, touche à sa fin, prédit Thomas Germain pour la BBC. Nous entrons dans « le web des machines », le web synthétique, le web artificiel où tous les contenus sont appelés à être générés en permanence, à la volée, en s’appuyant sur l’ensemble des contenus disponibles, sans que ceux-ci soient encore disponibles voire accessibles. Un second web vient se superposer au premier, le recouvrir… avec le risque de faire disparaître le web que nous avons connu, construit, façonné. 

Jusqu’à présent, le web reposait sur un marché simple, rappelle Germain. Les sites laissaient les moteurs de recherche indexer leurs contenus et les moteurs de recherche redirigeaient les internautes vers les sites web référencés. « On estime que 68 % de l’activité Internet commence sur les moteurs de recherche et qu’environ 90 % des recherches se font sur Google. Si Internet est un jardin, Google est le soleil qui fait pousser les fleurs »

Ce système a été celui que nous avons connu depuis les origines du web. L’intégration de l’IA, pour le meilleur ou pour le pire, promet néanmoins de transformer radicalement cette expérience. Confronté à une nette dégradation des résultats de la recherche, notamment due à l’affiliation publicitaire et au spam, le PDG de Google, Sundar Pichai, a promis une « réinvention totale de la recherche » en lançant son nouveau « mode IA ». Contrairement aux aperçus IA disponibles jusqu’à présent, le mode IA va remplacer complètement les résultats de recherche traditionnels. Désormais, un chatbot va créer un article pour répondre aux questions. En cours de déploiement et facultatif pour l’instant, à terme, il sera « l’avenir de la recherche Google »

Un détournement massif de trafic

Les critiques ont montré que, les aperçus IA généraient déjà beaucoup moins de trafic vers le reste d’internet (de 30 % à 70 %, selon le type de recherche. Des analyses ont également révélé qu’environ 60 % des recherches Google depuis le lancement des aperçus sont désormais « zéro clic », se terminant sans que l’utilisateur ne clique sur un seul lien – voir les études respectives de SeerInteractive, Semrush, Bain et Sparktoro), et beaucoup craignent que le mode IA ne renforce encore cette tendance. Si cela se concrétise, cela pourrait anéantir le modèle économique du web tel que nous le connaissons. Google estime que ces inquiétudes sont exagérées, affirmant que le mode IA « rendra le web plus sain et plus utile ». L’IA dirigerait les utilisateurs vers « une plus grande diversité de sites web » et le trafic serait de « meilleure qualité » car les utilisateurs passent plus de temps sur les liens sur lesquels ils cliquent. Mais l’entreprise n’a fourni aucune donnée pour étayer ces affirmations. 

Google et ses détracteurs s’accordent cependant sur un point : internet est sur le point de prendre une toute autre tournure. C’est le principe même du web qui est menacé, celui où chacun peut créer un site librement accessible et référencé. 

L’article de la BBC remarque, très pertinemment, que cette menace de la mort du web a déjà été faite. En 2010, Wired annonçait « la mort du web ». A l’époque, l’essor des smartphones, des applications et des réseaux sociaux avaient déjà suscité des prédictions apocalyptiques qui ne se sont pas réalisées. Cela n’empêche pas les experts d’être soucieux face aux transformations qui s’annoncent. Pour les critiques, certes, les aperçus IA et le mode IA incluent tous deux des liens vers des sources, mais comme l’IA vous donne la réponse que vous cherchez, cliquer sur ceux-ci devient superflu. C’est comme demander un livre à un bibliothécaire et qu’il vous en parle plutôt que de vous le fournir, compare un expert. 

La chute du nombre de visiteurs annoncée pourrait faire la différence entre une entreprise d’édition viable… et la faillite. Pour beaucoup d’éditeurs, ce changement sera dramatique. Nombre d’entreprises constatent que Google affiche leurs liens plus souvent, mais que ceux-ci sont moins cliqués. Selon le cabinet d’analyse de données BrightEdge, les aperçus IA ont entraîné une augmentation de 49 % des impressions sur le web, mais les clics ont chuté de 30 %, car les utilisateurs obtiennent leurs réponses directement de l’IA. « Google a écrit les règles, créé le jeu et récompensé les joueurs », explique l’une des expertes interrogée par la BBC. « Maintenant, ils se retournent et disent : « C’est mon infrastructure, et le web se trouve juste dedans ». »

Demis Hassabis, directeur de Google DeepMind, le laboratoire de recherche en IA de l’entreprise, a déclaré qu’il pensait que demain, les éditeurs alimenteraient directement les modèles d’IA avec leurs contenus, sans plus avoir à se donner la peine de publier des informations sur des sites web accessibles aux humains. Mais, pour Matthew Prince, directeur général de Cloudflare, le problème dans ce web automatisé, c’est que « les robots ne cliquent pas sur les publicités ». « Si l’IA devient l’audience, comment les créateurs seront-ils rémunérés ? » La rémunération directe existe déjà, comme le montrent les licences de contenus que les plus grands éditeurs de presse négocient avec des systèmes d’IA pour qu’elles s’entraînent et exploitent leurs contenus, mais ces revenus là ne compenseront pas la chute d’audience à venir. Et ce modèle ne passera certainement pas l’échelle d’une rétribution généralisée. 

Si gagner de l’argent sur le web devient plus difficile, il est probable que nombre d’acteurs se tournent vers les réseaux sociaux pour tenter de compenser les pertes de revenus. Mais là aussi, les caprices algorithmiques et le développement de l’IA générative risquent de ne pas suffire à compenser les pertes. 

Un nouvel internet sans condition

Pour Google, les réactions aux aperçus IA laissent présager que le mode IA sera extrêmement populaire. « À mesure que les utilisateurs utilisent AI Overviews, nous constatons qu’ils sont plus satisfaits de leurs résultats et effectuent des recherches plus souvent », a déclaré Pichai lors de la conférence des développeurs de Google. Autrement dit, Google affirme que cela améliore la recherche et que c’est ce que veulent les utilisateurs. Mais pour Danielle Coffey, présidente de News/Media Alliance, un groupement professionnel représentant plus de 2 200 journalistes et médias, les réponses de l’IA vont remplacer les produits originaux : « les acteurs comme Google vont gagner de l’argent grâce à notre contenu et nous ne recevons rien en retour ». Le problème, c’est que Google n’a pas laissé beaucoup de choix aux éditeurs, comme le pointait Bloomberg. Soit Google vous indexe pour la recherche et peut utiliser les contenus pour ses IA, soit vous êtes désindexé des deux. La recherche est bien souvent l’une des premières utilisations de outils d’IA. Les inquiétudes sur les hallucinations, sur le renforcement des chambres d’échos dans les réponses que vont produire ces outils sont fortes (on parle même de « chambre de chat » pour évoquer la réverbération des mêmes idées et liens dans ces outils). Pour Cory Doctorow, « Google s’apprête à faire quelque chose qui va vraiment mettre les gens en colère »… et appelle les acteurs à capitaliser sur cette colère à venir. Matthew Prince de Cloudflare prône, lui, une intervention directe. Son projet est de faire en sorte que Cloudflare et un consortium d’éditeurs de toutes tailles bloquent collectivement les robots d’indexation IA, à moins que les entreprises technologiques ne paient pour le contenu. Il s’agit d’une tentative pour forcer la Silicon Valley à négocier. « Ma version très optimiste », explique Prince, « est celle où les humains obtiennent du contenu gratuitement et où les robots doivent payer une fortune pour l’obtenir ». Tim O’Reilly avait proposé l’année dernière quelque chose d’assez similaire : expliquant que les droits dérivés liés à l’exploitation des contenus par l’IA devraient donner lieu à rétribution – mais à nouveau, une rétribution qui restera par nature insuffisante, comme l’expliquait Frédéric Fillioux

Même constat pour le Washington Post, qui s’inquiète de l’effondrement de l’audience des sites d’actualité avec le déploiement des outils d’IA. « Le trafic de recherche organique vers ses sites web a diminué de 55 % entre avril 2022 et avril 2025, selon les données de Similarweb ». Dans la presse américaine, l’audience est en berne et les licenciements continuent.

Les erreurs seront dans la réponse

Pour la Technology Review, c’est la fin de la recherche par mots-clés et du tri des liens proposés. « Nous entrons dans l’ère de la recherche conversationnelle » dont la fonction même vise à « ignorer les liens », comme l’affirme Perplexity dans sa FAQ. La TR rappelle l’histoire de la recherche en ligne pour montrer que des annuaires aux moteurs de recherche, celle-ci a toujours proposé des améliorations, pour la rendre plus pertinente. Depuis 25 ans, Google domine la recherche en ligne et n’a cessé de s’améliorer pour fournir de meilleures réponses. Mais ce qui s’apprête à changer avec l’intégration de l’IA, c’est que les sources ne sont plus nécessairement accessibles et que les réponses sont générées à la volée, aucune n’étant identique à une autre. 

L’intégration de l’IA pose également la question de la fiabilité des réponses. L’IA de Google a par exemple expliqué que la Technology Review avait été mise en ligne en 2022… ce qui est bien sûr totalement faux, mais qu’en saurait une personne qui ne le sait pas ? Mais surtout, cet avenir génératif promet avant tout de fabriquer des réponses à la demande. Mat Honan de la TR donne un exemple : « Imaginons que je veuille voir une vidéo expliquant comment réparer un élément de mon vélo. La vidéo n’existe pas, mais l’information, elle, existe. La recherche générative assistée par l’IA pourrait théoriquement trouver cette information en ligne – dans un manuel d’utilisation caché sur le site web d’une entreprise, par exemple – et créer une vidéo pour me montrer exactement comment faire ce que je veux, tout comme elle pourrait me l’expliquer avec des mots aujourd’hui » – voire très mal nous l’expliquer. L’exemple permet de comprendre comment ce nouvel internet génératif pourrait se composer à la demande, quelque soit ses défaillances. 

Mêmes constats pour Matteo Wrong dans The Atlantic : avec la généralisation de l’IA, nous retournons dans un internet en mode bêta. Les services et produits numériques n’ont jamais été parfaits, rappelle-t-il, mais la généralisation de l’IA risque surtout d’amplifier les problèmes. Les chatbots sont très efficaces pour produire des textes convaincants, mais ils ne prennent pas de décisions en fonction de l’exactitude factuelle. Les erreurs sont en passe de devenir « une des caractéristiques de l’internet ». « La Silicon Valley mise l’avenir du web sur une technologie capable de dérailler de manière inattendue, de s’effondrer à la moindre tâche et d’être mal utilisée avec un minimum de frictions ». Les quelques réussites de l’IA n’ont que peu de rapport avec la façon dont de nombreuses personnes et entreprises comprennent et utilisent cette technologie, rappelle-t-il. Plutôt que des utilisations ciblées et prudentes, nombreux sont ceux qui utilisent l’IA générative pour toutes les tâches imaginables, encouragés par les géants de la tech. « Tout le monde utilise l’IA pour tout », titrait le New York Times. « C’est là que réside le problème : l’IA générative est une technologie suffisamment performante pour que les utilisateurs en deviennent dépendants, mais pas suffisamment fiable pour être véritablement fiable ». Nous allons vers un internet où chaque recherche, itinéraire, recommandation de restaurant, résumé d’événement, résumé de messagerie vocale et e-mail sera plus suspect qu’il n’est aujourd’hui. « Les erreurs d’aujourd’hui pourraient bien, demain, devenir la norme », rendant ses utilisateurs incapables de vérifier ses fonctionnements. Bienvenue dans « l’âge de la paranoïa », clame Wired.

Vers la publicité générative et au-delà !

Mais il n’y a pas que les « contenus » qui vont se recomposer, la publicité également. C’est ainsi qu’il faut entendre les déclarations de Mark Zuckerberg pour automatiser la création publicitaire, explique le Wall Street Journal. « La plateforme publicitaire de Meta propose déjà des outils d’IA capables de générer des variantes de publicités existantes et d’y apporter des modifications mineures avant de les diffuser aux utilisateurs sur Facebook et Instagram. L’entreprise souhaite désormais aider les marques à créer des concepts publicitaires de A à Z ». La publicité représente 97% du chiffre d’affaires de Meta, rappelle le journal (qui s’élève en 2024 à 164 milliards de dollars). Chez Meta les contenus génératifs produisent déjà ce qu’on attend d’eux. Meta a annoncé une augmentation de 8 % du temps passé sur Facebook et de 6 % du temps passé sur Instagram grâce aux contenus génératifs. 15 millions de publicités par mois sur les plateformes de Meta sont déjà générées automatiquement. « Grâce aux outils publicitaires développés par Meta, une marque pourrait demain fournir une image du produit qu’elle souhaite promouvoir, accompagnée d’un objectif budgétaire. L’IA créerait alors l’intégralité de la publicité, y compris les images, la vidéo et le texte. Le système déciderait ensuite quels utilisateurs Instagram et Facebook cibler et proposerait des suggestions en fonction du budget ». Selon la géolocalisation des utilisateurs, la publicité pourrait s’adapter en contexte, créant l’image d’une voiture circulant dans la neige ou sur une plage s’ils vivent en montagne ou au bord de la mer. « Dans un avenir proche, nous souhaitons que chaque entreprise puisse nous indiquer son objectif, comme vendre quelque chose ou acquérir un nouveau client, le montant qu’elle est prête à payer pour chaque résultat, et connecter son compte bancaire ; nous nous occuperons du reste », a déclaré Zuckerberg lors de l’assemblée générale annuelle des actionnaires de l’entreprise. 

Nilay Patel, le rédac chef de The Verge, parle de « créativité infinie ». C’est d’ailleurs la même idée que l’on retrouve dans les propos de Jensen Huang, le PDG de Nvidia, quand il promet de fabriquer les « usines à IA » qui généreront le web demain. Si toutes les grandes entreprises et les agences de publicité ne sont pas ravies de la proposition – qui leur est fondamentalement hostile, puisqu’elle vient directement les concurrencer -, d’autres s’y engouffrent déjà, à l’image d’Unilever qui explique sur Adweek que l’IA divise par deux ses budgets publicitaires grâce à son partenariat avec Nvidia. « Unilever a déclaré avoir réalisé jusqu’à 55 % d’économies sur ses campagnes IA, d’avoir réduit les délais de production de 65% tout en doublant le taux de clic et en retenant l’attention des consommateurs trois fois plus longtemps »

L’idée finalement très partagée par tous les géants de l’IA, c’est bien d’annoncer le remplacement du web que l’on connaît par un autre. Une sous-couche générative qu’il maîtriseraient, capable de produire un web à leur profit, qu’ils auraient avalé et digéré. 

Vers des revenus génératifs ?

Nilay Patel était l’année dernière l’invité du podcast d’Ezra Klein pour le New York Times qui se demandait si cette transformation du web allait le détruire ou le sauver. Dans cette discussion parfois un peu décousue, Klein rappelle que l’IA se développe d’abord là où les produits n’ont pas besoin d’être très performants. Des tâches de codage de bas niveau aux devoirs des étudiants, il est également très utilisé pour la diffusion de contenus médiocres sur l’internet. Beaucoup des contenus d’internet ne sont pas très performants, rappelle-t-il. Du spam au marketing en passant par les outils de recommandations des réseaux sociaux, internet est surtout un ensemble de contenus à indexer pour délivrer de la publicité elle-même bien peu performante. Et pour remplir cet « internet de vide », l’IA est assez efficace. Les plateformes sont désormais inondées de contenus sans intérêts, de spams, de slops, de contenus de remplissage à la recherche de revenus. Et Klein de se demander que se passera-t-il lorsque ces flots de contenu IA s’amélioreront ? Que se passera-t-il lorsque nous ne saurons plus s’il y a quelqu’un à l’autre bout du fil de ce que nous voyons, lisons ou entendons ? Y aura-t-il encore quelqu’un d’ailleurs, où n’aurons nous accès plus qu’à des contenus génératifs ?

Pour Patel, pour l’instant, l’IA inonde le web de contenus qui le détruisent. En augmentant à l’infini l’offre de contenu, le système s’apprête à s’effondrer sur lui-même : « Les algorithmes de recommandation s’effondrent, notre capacité à distinguer le vrai du faux s’effondre également, et, plus important encore, les modèles économiques d’Internet s’effondrent complètement ». Les contenus n’arrivent plus à trouver leurs publics, et inversement. L’exemple éclairant pour illustrer cela, c’est celui d’Amazon. Face à l’afflux de livres générés par l’IA, la seule réponse d’Amazon a été de limiter le nombre de livres déposables sur la plateforme à trois par jour. C’est une réponse parfaitement absurde qui montre que nos systèmes ne sont plus conçus pour organiser leurs publics et leur adresser les bons contenus. C’est à peine s’ils savent restreindre le flot

Avec l’IA générative, l’offre ne va pas cesser d’augmenter. Elle dépasse déjà ce que nous sommes capables d’absorber individuellement. Pas étonnant alors que toutes les plateformes se transforment de la même manière en devenant des plateformes de téléachats ne proposant plus rien d’autre que de courtes vidéos.

« Toutes les plateformes tendent vers le même objectif, puisqu’elles sont soumises aux mêmes pressions économiques ». Le produit des plateformes c’est la pub. Elles mêmes ne vendent rien. Ce sont des régies publicitaires que l’IA promet d’optimiser depuis les données personnelles collectées. Et demain, nos boîtes mails seront submergées de propositions marketing générées par l’IA… Pour Patel, les géants du net ont arrêté de faire leur travail. Aucun d’entre eux ne nous signale plus que les contenus qu’ils nous proposent sont des publicités. Google Actualités référence des articles écrits par des IA sans que cela ne soit un critère discriminant pour les référenceurs de Google, expliquait 404 média (voir également l’enquête de Next sur ce sujet qui montre que les sites générés par IA se démultiplient, « pour faire du fric »). Pour toute la chaîne, les revenus semblent être devenus le seul objectif.

Et Klein de suggérer que ces contenus vont certainement s’améliorer, comme la génération d’image et de texte n’a cessé de s’améliorer. Il est probable que l’article moyen d’ici trois ans sera meilleur que le contenu moyen produit par un humain aujourd’hui. « Je me suis vraiment rendu compte que je ne savais pas comment répondre à la question : est-ce un meilleur ou un pire internet qui s’annonce ? Pour répondre presque avec le point de vue de Google, est-ce important finalement que le contenu soit généré par un humain ou une IA, ou est-ce une sorte de sentimentalisme nostalgique de ma part ? » 

Il y en a certainement, répond Patel. Il n’y a certainement pas besoin d’aller sur une page web pour savoir combien de temps il faut pour cuire un œuf, l’IA de Google peut vous le dire… Mais, c’est oublier que cette IA générative ne sera pas plus neutre que les résultats de Google aujourd’hui. Elle sera elle aussi façonnée par la publicité. L’enjeu demain ne sera plus d’être dans les 3 premiers résultats d’une page de recherche, mais d’être citée par les réponses construites par les modèles de langages. « Votre client le plus important, désormais, c’est l’IA ! », explique le journaliste Scott Mulligan pour la Technology Review. « L’objectif ultime n’est pas seulement de comprendre comment votre marque est perçue par l’IA, mais de modifier cette perception ». Or, les biais marketing des LLM sont déjà nombreux. Une étude montre que les marques internationales sont souvent perçues comme étant de meilleures qualités que les marques locales. Si vous demandez à un chatbot de recommander des cadeaux aux personnes vivant dans des pays à revenu élevé, il suggérera des articles de marque de luxe, tandis que si vous lui demandez quoi offrir aux personnes vivant dans des pays à faible revenu, il recommandera des marques plus cheap.

L’IA s’annonce comme un nouveau public des marques, à dompter. Et la perception d’une marque par les IA aura certainement des impacts sur leurs résultats financiers. Le marketing a assurément trouvé un nouveau produit à vendre ! Les entreprises vont adorer !

Pour Klein, l’internet actuel est certes très affaibli, pollué de spams et de contenus sans intérêts. Google, Meta et Amazon n’ont pas créé un internet que les gens apprécient, mais bien plus un internet que les gens utilisent à leur profit. L’IA propose certainement non pas un internet que les gens vont plus apprécier, bien au contraire, mais un internet qui profite aux grands acteurs plutôt qu’aux utilisateurs. Pour Patel, il est possible qu’un internet sans IA subsiste, pour autant qu’il parvienne à se financer.

Pourra-t-on encore défendre le web que nous voulons ?

Les acteurs oligopolistiques du numérique devenus les acteurs oligopolistiques de l’IA semblent s’aligner pour transformer le web à leur seul profit, et c’est assurément la puissance (et surtout la puissance financière) qu’ils ont acquis qui le leur permet. La transformation du web en « web des machines » est assurément la conséquence de « notre longue dépossession », qu’évoquait Ben Tarnoff dans son livre, Internet for the People.

La promesse du web synthétique est là pour rester. Et la perspective qui se dessine, c’est que nous avons à nous y adapter, sans discussion. Ce n’est pas une situation très stimulante, bien au contraire. A mesure que les géants de l’IA conquièrent le numérique, c’est nos marges de manœuvres qui se réduisent. Ce sont elles que la régulation devrait chercher à réouvrir, dès à présent. Par exemple en mobilisant très tôt le droit à la concurrence et à l’interopérabilité, pour forcer les acteurs à proposer aux utilisateurs d’utiliser les IA de leurs choix ou en leur permettant, très facilement, de refuser leur implémentations dans les outils qu’ils utilisent, que ce soit leurs OS comme les services qu’ils utilisent. Bref, mobiliser le droit à la concurrence et à l’interopérabilité au plus tôt. Afin que défendre le web que nous voulons ne s’avère pas plus difficile demain qu’il n’était aujourd’hui.

Hubert Guillaud

Cet édito a été originellement publié dans la première lettre d’information de CaféIA le 27 juin 2025.

  • ✇404 Media
  • ICE Is Searching a Massive Insurance and Medical Bill Database to Find Deportation Targets
    Agents from Immigration and Customs Enforcement (ICE) have gained access to a massive database of health and car insurance claims and are using it to track down people they want to deport, according to internal ICE material viewed by 404 Media. The database, which contains details on more than 1.8 billion insurance claims and 58 million medical bills and growing, includes peoples’ names, addresses, telephone and tax identification numbers, license plates, and other sensitive personal informat
     

ICE Is Searching a Massive Insurance and Medical Bill Database to Find Deportation Targets

9 juillet 2025 à 09:38
ICE Is Searching a Massive Insurance and Medical Bill Database to Find Deportation Targets

Agents from Immigration and Customs Enforcement (ICE) have gained access to a massive database of health and car insurance claims and are using it to track down people they want to deport, according to internal ICE material viewed by 404 Media. The database, which contains details on more than 1.8 billion insurance claims and 58 million medical bills and growing, includes peoples’ names, addresses, telephone and tax identification numbers, license plates, and other sensitive personal information.

The news shows how ICE continues to try to leverage whatever data it is able to access or purchase as part of its deportation mission. The news also highlights the existence of the database, called ISO ClaimSearch, that many members of the public have likely never heard of, nor understand they may be included in. Traditionally ISO ClaimSearch is used by insurers to identify people committing fraud or police to recover stolen vehicles. Now, that database is being repurposed as a deportation tool.

💡
Do you know anything else about ICE's access to datasets like this? I would love to hear from you. Using a non-work device, you can message me securely on Signal at joseph.404 or send me an email at joseph@404media.co.

“ICE ERO use of this data reaffirms that ICE will stop at nothing to build a mass surveillance dragnet to track, surveil and criminalize all community members. Time and time again, ICE has shown us that it intends to build a mass surveillance system that nets all Americans. It is not about combatting crime, this is about the federal government having surveillance power and control over all Americans,” Julie Mao, co-founder and deputy director of Just Futures Law, told 404 Media in an email.

  • ✇404 Media
  • Podcast: How to Fight Back Against AI Bot Scrapers
    We’re back! We start this week with Emanuel’s article about Anubis, an open source piece of software that is saving the internet from AI bot scrapers. After the break, Joseph tells us about the new facial recognition app ICE is using and which he revealed. In the subscribers-only section, we do a lightning round runthrough of a bunch of our recent stories about LLMs and how to trick them, or what they don’t understand. Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Be
     

Podcast: How to Fight Back Against AI Bot Scrapers

9 juillet 2025 à 09:00
Podcast: How to Fight Back Against AI Bot Scrapers

We’re back! We start this week with Emanuel’s article about Anubis, an open source piece of software that is saving the internet from AI bot scrapers. After the break, Joseph tells us about the new facial recognition app ICE is using and which he revealed. In the subscribers-only section, we do a lightning round runthrough of a bunch of our recent stories about LLMs and how to trick them, or what they don’t understand.

Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a paid subscriber, check your inbox for an email from our podcast host Transistor for a link to the subscribers-only version! You can also add that subscribers feed to your podcast app of choice and never miss an episode that way. The email should also contain the subscribers-only unlisted YouTube link for the extended video version too. It will also be in the show notes in your podcast player.

  • ✇404 Media
  • 'Save Our Signs' Wants to Save the Real History of National Parks Before Trump Erases It
    Data preservationists and archivists have been working tirelessly since the election of President Donald Trump to save websites, data, and public information that’s being removed by the administration for promoting or even mentioning diversity. The administration is now targeting National Parks signs that educate visitors about anything other than “beauty” and “grandeur,” and demanding they remove signs that mention “negative” aspects of American history. In March, Trump issued an executive o
     

'Save Our Signs' Wants to Save the Real History of National Parks Before Trump Erases It

8 juillet 2025 à 12:22
'Save Our Signs' Wants to Save the Real History of National Parks Before Trump Erases It

Data preservationists and archivists have been working tirelessly since the election of President Donald Trump to save websites, data, and public information that’s being removed by the administration for promoting or even mentioning diversity. The administration is now targeting National Parks signs that educate visitors about anything other than “beauty” and “grandeur,” and demanding they remove signs that mention “negative” aspects of American history. 

In March, Trump issued an executive order, titled “Restoring Truth and Sanity To American History,” demanding public officials ensure that public monuments and markers under the Department of the Interior’s jurisdiction never address anything negative about American history, past or present. Instead, Trump wrote, they should only ever acknowledge how pretty the landscape looks.

💡
Do you know anything else about how the Trump administration is affecting the National Park Service? I would love to hear from you. Using a non-work device, you can message me securely on Signal at sam.404. Otherwise, send me an email at sam@404media.co.

Last month, National Park Service directors across the country were informed that they must post surveys at informational sites that encourage visitors to report "any signs or other information that are negative about either past or living Americans or that fail to emphasize the beauty, grandeur, and abundance of landscapes and other natural features," as dictated in a May follow-up order from Interior Secretary Doug Burgum. QR codes started popping up on placards in national parks that take visitors to a survey that asks them to snitch on "any signs or other information that are negative about either past or living Americans or that fail to emphasize the beauty, grandeur, and abundance of landscapes and other natural features." 

  • ✇404 Media
  • Researchers Jailbreak AI by Flooding It With Bullshit Jargon
    You can trick AI chatbots like ChatGPT or Gemini into teaching you how to make a bomb or hack an ATM if you make the question complicated, full of academic jargon, and cite sources that do not exist. That’s the conclusion of a new paper authored by a team of researchers from Intel, Boise State University, and University of Illinois at Urbana-Champaign. The research details this new method of jailbreaking LLMs, called “Information Overload” by the researchers, and an automated system for attack t
     

Researchers Jailbreak AI by Flooding It With Bullshit Jargon

8 juillet 2025 à 09:00
Researchers Jailbreak AI by Flooding It With Bullshit Jargon

You can trick AI chatbots like ChatGPT or Gemini into teaching you how to make a bomb or hack an ATM if you make the question complicated, full of academic jargon, and cite sources that do not exist. 

That’s the conclusion of a new paper authored by a team of researchers from Intel, Boise State University, and University of Illinois at Urbana-Champaign. The research details this new method of jailbreaking LLMs, called “Information Overload” by the researchers, and an automated system for attack they call “InfoFlood.” The paper, titled “InfoFlood: Jailbreaking Large Language Models with Information Overload” was published as a preprint.

Popular LLMs like ChatGPT, Gemini, or LLaMA have guardrails that stop them from answering some questions. ChatGPT will not, for example, tell you how to build a bomb or talk someone into suicide if you ask it in a straightforward manner. But people can “jailbreak” LLMs by asking questions the right way and circumvent those protections.

  • ✇Dans les algorithmes
  • Renverser le pouvoir artificiel
    L’AI Now Institute vient de publier son rapport 2025. Et autant dire, qu’il frappe fort. “La trajectoire actuelle de l’IA ouvre la voie à un avenir économique et politique peu enviable : un avenir qui prive de leurs droits une grande partie du public, rend les systèmes plus obscurs pour ceux qu’ils affectent, dévalorise notre savoir-faire, compromet notre sécurité et restreint nos perspectives d’innovation”.  La bonne nouvelle, c’est que la voie offerte par l’industrie technologique n’est pas
     

Renverser le pouvoir artificiel

8 juillet 2025 à 01:00

L’AI Now Institute vient de publier son rapport 2025. Et autant dire, qu’il frappe fort. “La trajectoire actuelle de l’IA ouvre la voie à un avenir économique et politique peu enviable : un avenir qui prive de leurs droits une grande partie du public, rend les systèmes plus obscurs pour ceux qu’ils affectent, dévalorise notre savoir-faire, compromet notre sécurité et restreint nos perspectives d’innovation”

La bonne nouvelle, c’est que la voie offerte par l’industrie technologique n’est pas la seule qui s’offre à nous. “Ce rapport explique pourquoi la lutte contre la vision de l’IA défendue par l’industrie est un combat qui en vaut la peine”. Comme le rappelait leur rapport 2023, l’IA est d’abord une question de concentration du pouvoir entre les mains de quelques géants. “La question que nous devrions nous poser n’est pas de savoir si ChatGPT est utile ou non, mais si le pouvoir irréfléchi d’OpenAI, lié au monopole de Microsoft et au modèle économique de l’économie technologique, est bénéfique à la société”

“L’avènement de ChatGPT en 2023 ne marque pas tant une rupture nette dans l’histoire de l’IA, mais plutôt le renforcement d’un paradigme du « plus c’est grand, mieux c’est », ancré dans la perpétuation des intérêts des entreprises qui ont bénéficié du laxisme réglementaire et des faibles taux d’intérêt de la Silicon Valley”. Mais ce pouvoir ne leur suffit pas : du démantèlement des gouvernements au pillage des données, de la dévalorisation du travail pour le rendre compatible à l’IA, à la réorientation des infrastructures énergétiques en passant par le saccage de l’information et de la démocratie… l’avènement de l’IA exige le démantèlement de nos infrastructures sociales, politiques et économiques au profit des entreprises de l’IA. L’IA remet au goût du jour des stratégies anciennes d’extraction d’expertises et de valeurs pour concentrer le pouvoir entre les mains des extracteurs au profit du développement de leurs empires. 

Mais pourquoi la société accepterait-elle un tel compromis, une telle remise en cause ? Pour les chercheurs.ses de l’AI Now Institute ce pouvoir doit et peut être perturbé, notamment parce qu’il est plus fragile qu’il n’y paraît. “Les entreprises d’IA perdent de l’argent pour chaque utilisateur qu’elles gagnent” et le coût de l’IA à grande échelle va être très élevé au risque qu’une bulle d’investissement ne finisse par éclater. L’affirmation de la révolution de l’IA générative, elle, contraste avec la grande banalité de ses intégrations et les difficultés qu’elle engendre : de la publicité automatisée chez Meta, à la production de code via Copilot (au détriment des compétences des développeurs), ou via la production d’agents IA, en passant par l’augmentation des prix du Cloud par l’intégration automatique de fonctionnalités IA… tout en laissant les clients se débrouiller des hallucinations, des erreurs et des imperfactions de leurs produits. Or, appliqués en contexte réel les systèmes d’IA échouent profondément même sur des tâches élémentaires, rappellent les auteurs du rapport : les fonctionnalités de l’IA relèvent souvent d’illusions sur leur efficacité, masquant bien plus leurs défaillances qu’autre chose, comme l’expliquent les chercheurs Inioluwa Deborah Raji, Elizabeth Kumar, Aaron Horowitz et Andrew D. Selbst. Dans de nombreux cas d’utilisation, “l’IA est déployée par ceux qui ont le pouvoir contre ceux qui n’en ont pas” sans possibilité de se retirer ou de demander réparation en cas d’erreur.

L’IA : un outil défaillant au service de ceux qui la déploie

Pour l’AI Now Institute, les avantages de l’IA sont à la fois surestimés et sous-estimés, des traitements contre le cancer à une hypothétique croissance économique, tandis que certains de ses défauts sont réels, immédiats et se répandent. Le solutionnisme de l’IA occulte les problèmes systémiques auxquels nos économies sont confrontées, occultant la concentration économique à l’oeuvre et servant de canal pour le déploiement de mesures d’austérité sous prétexte d’efficacité, à l’image du très problématique chatbot mis en place par la ville New York. Des millions de dollars d’argent public ont été investis dans des solutions d’IA défaillantes. “Le mythe de la productivité occulte une vérité fondamentale : les avantages de l’IA profitent aux entreprises, et non aux travailleurs ou au grand public. Et L’IA agentive rendra les lieux de travail encore plus bureaucratiques et surveillés, réduisant l’autonomie au lieu de l’accroître”. 

“L’utilisation de l’IA est souvent coercitive”, violant les droits et compromettant les procédures régulières à l’image de l’essor débridé de l’utilisation de l’IA dans le contrôle de l’immigration aux Etats-Unis (voir notre article sur la fin du cloisonnement des données ainsi que celui sur l’IA générative, nouvelle couche d’exploitation du travail). Le rapport consacre d’ailleurs tout un chapitre aux défaillances de l’IA. Pour les thuriféraires de l’IA, celle-ci est appelée à guérir tous nos maux, permettant à la fois de transformer la science, la logistique, l’éducation… Mais, si les géants de la tech veulent que l’IA soit accessible à tous, alors l’IA devrait pouvoir bénéficier à tous. C’est loin d’être le cas. 

Le rapport prend l’exemple de la promesse que l’IA pourrait parvenir, à terme, à guérir les cancers. Si l’IA a bien le potentiel de contribuer aux recherches dans le domaine, notamment en améliorant le dépistage, la détection et le diagnostic. Il est probable cependant que loin d’être une révolution, les améliorations soient bien plus incrémentales qu’on le pense. Mais ce qui est contestable dans ce tableau, estiment les chercheurs de l’AI Now Institute, c’est l’hypothèse selon laquelle ces avancées scientifiques nécessitent la croissance effrénée des hyperscalers du secteur de l’IA. Or, c’est précisément le lien que ces dirigeants d’entreprise tentent d’établir. « Le prétexte que l’IA pourrait révolutionner la santé sert à promouvoir la déréglementation de l’IA pour dynamiser son développement ». Les perspectives scientifiques montées en promesses inéluctables sont utilisées pour abattre les résistances à discuter des enjeux de l’IA et des transformations qu’elle produit sur la société toute entière.

Or, dans le régime des défaillances de l’IA, bien peu de leurs promesses relèvent de preuves scientifiques. Nombre de recherches du secteur s’appuient sur un régime de véritude comme s’en moque l’humoriste Stephen Colbert, c’est-à-dire sur des recherches qui ne sont pas validées par les pairs, à l’image des robots infirmiers qu’a pu promouvoir Nvidia en affirmant qu’ils surpasseraient les infirmières elles-mêmes… Une affirmation qui ne reposait que sur une étude de Nvidia. Nous manquons d’une science de l’évaluation de l’IA générative. En l’absence de benchmarks indépendants et largement reconnus pour mesurer des attributs clés tels que la précision ou la qualité des réponses, les entreprises inventent leurs propres benchmarks et, dans certains cas, vendent à la fois le produit et les plateformes de validation des benchmarks au même client. Par exemple, Scale AI détient des contrats de plusieurs centaines de millions de dollars avec le Pentagone pour la production de modèles d’IA destinés au déploiement militaire, dont un contrat de 20 millions de dollars pour la plateforme qui servira à évaluer la précision des modèles d’IA destinés aux agences de défense. Fournir la solution et son évaluation est effectivement bien plus simple. 

Autre défaillance systémique : partout, les outils marginalisent les professionnels. Dans l’éducation, les Moocs ont promis la démocratisation de l’accès aux cours. Il n’en a rien été. Désormais, le technosolutionnisme promet la démocratisation par l’IA générative via des offres dédiées comme ChatGPT Edu d’OpenAI, au risque de compromettre la finalité même de l’éducation. En fait, rappellent les auteurs du rapport, dans l’éducation comme ailleurs, l’IA est bien souvent adoptée par des administrateurs, sans discussion ni implication des concernés. A l’université, les administrateurs achètent des solutions non éprouvées et non testées pour des sommes considérables afin de supplanter les technologies existantes gérées par les services technologiques universitaires. Même constat dans ses déploiements au travail, où les pénuries de main d’œuvre sont souvent évoquées comme une raison pour développer l’IA, alors que le problème n’est pas tant la pénurie que le manque de protection ou le régime austéritaire de bas salaires. Les solutions technologiques permettent surtout de rediriger les financements au détriment des travailleurs et des bénéficiaires. L’IA sert souvent de vecteur pour le déploiement de mesures d’austérité sous un autre nom. Les systèmes d’IA appliqués aux personnes à faibles revenus n’améliorent presque jamais l’accès aux prestations sociales ou à d’autres opportunités, disait le rapport de Techtonic Justice. “L’IA n’est pas un ensemble cohérent de technologies capables d’atteindre des objectifs sociaux complexes”. Elle est son exact inverse, explique le rapport en pointant par exemple les défaillances du Doge (que nous avons nous-mêmes documentés). Cela n’empêche pourtant pas le solutionnisme de prospérer. L’objectif du chatbot newyorkais par exemple, “n’est peut-être pas, en réalité, de servir les citoyens, mais plutôt d’encourager et de centraliser l’accès aux données des citoyens ; de privatiser et d’externaliser les tâches gouvernementales ; et de consolider le pouvoir des entreprises sans mécanismes de responsabilisation significatifs”, comme l’explique le travail du Surveillance resistance Lab, très opposé au projet.

Le mythe de la productivité enfin, que répètent et anônnent les développeurs d’IA, nous fait oublier que les bénéfices de l’IA vont bien plus leur profiter à eux qu’au public. « La productivité est un euphémisme pour désigner la relation économique mutuellement bénéfique entre les entreprises et leurs actionnaires, et non entre les entreprises et leurs salariés. Non seulement les salariés ne bénéficient pas des gains de productivité liés à l’IA, mais pour beaucoup, leurs conditions de travail vont surtout empirer. L’IA ne bénéficie pas aux salariés, mais dégrade leurs conditions de travail, en augmentant la surveillance, notamment via des scores de productivité individuels et collectifs. Les entreprises utilisent la logique des gains de productivité de l’IA pour justifier la fragmentation, l’automatisation et, dans certains cas, la suppression du travail. » Or, la logique selon laquelle la productivité des entreprises mènera inévitablement à une prospérité partagée est profondément erronée. Par le passé, lorsque l’automatisation a permis des gains de productivité et des salaires plus élevés, ce n’était pas grâce aux capacités intrinsèques de la technologie, mais parce que les politiques des entreprises et les réglementations étaient conçues de concert pour soutenir les travailleurs et limiter leur pouvoir, comme l’expliquent Daron Acemoglu et Simon Johnson, dans Pouvoir et progrès (Pearson 2024). L’essor de l’automatisation des machines-outils autour de la Seconde Guerre mondiale est instructif : malgré les craintes de pertes d’emplois, les politiques fédérales et le renforcement du mouvement ouvrier ont protégé les intérêts des travailleurs et exigé des salaires plus élevés pour les ouvriers utilisant les nouvelles machines. Les entreprises ont à leur tour mis en place des politiques pour fidéliser les travailleurs, comme la redistribution des bénéfices et la formation, afin de réduire les turbulences et éviter les grèves. « Malgré l’automatisation croissante pendant cette période, la part des travailleurs dans le revenu national est restée stable, les salaires moyens ont augmenté et la demande de travailleurs a augmenté. Ces gains ont été annulés par les politiques de l’ère Reagan, qui ont donné la priorité aux intérêts des actionnaires, utilisé les menaces commerciales pour déprécier les normes du travail et les normes réglementaires, et affaibli les politiques pro-travailleurs et syndicales, ce qui a permis aux entreprises technologiques d’acquérir une domination du marché et un contrôle sur des ressources clés. L’industrie de l’IA est un produit décisif de cette histoire ». La discrimination salariale algorithmique optimise les salaires à la baisse. D’innombrables pratiques sont mobilisées pour isoler les salariés et contourner les lois en vigueur, comme le documente le rapport 2025 de FairWork. La promesse que les agents IA automatiseront les tâches routinières est devenue un point central du développement de produits, même si cela suppose que les entreprises qui s’y lancent deviennent plus processuelles et bureaucratiques pour leur permettre d’opérer. Enfin, nous interagissons de plus en plus fréquemment avec des technologies d’IA utilisées non pas par nous, mais sur nous, qui façonnent notre accès aux ressources dans des domaines allant de la finance à l’embauche en passant par le logement, et ce au détriment de la transparence et au détriment de la possibilité même de pouvoir faire autrement.

Le risque de l’IA partout est bien de nous soumettre aux calculs, plus que de nous en libérer. Par exemple, l’intégration de l’IA dans les agences chargées de l’immigration, malgré l’édiction de principes d’utilisation vertueux, montre combien ces principes sont profondément contournés, comme le montrait le rapport sur la déportation automatisée aux Etats-Unis du collectif de défense des droits des latino-américains, Mijente. Les Services de citoyenneté et d’immigration des États-Unis (USCIS) utilisent des outils prédictifs pour automatiser leurs prises de décision, comme « Asylum Text Analytics », qui interroge les demandes d’asile afin de déterminer celles qui sont frauduleuses. Ces outils ont démontré, entre autres défauts, des taux élevés d’erreurs de classification lorsqu’ils sont utilisés sur des personnes dont l’anglais n’est pas la langue maternelle. Les conséquences d’une identification erronée de fraude sont importantes : elles peuvent entraîner l’expulsion, l’interdiction à vie du territoire américain et une peine d’emprisonnement pouvant aller jusqu’à dix ans. « Pourtant, la transparence pour les personnes concernées par ces systèmes est plus que limitée, sans possibilité de se désinscrire ou de demander réparation lorsqu’ils sont utilisés pour prendre des décisions erronées, et, tout aussi important, peu de preuves attestent que l’efficacité de ces outils a été, ou peut être, améliorée »

Malgré la légalité douteuse et les failles connues de nombre de ces systèmes que le rapport documente, l’intégration de l’IA dans les contrôles d’immigration ne semble vouée qu’à s’intensifier. L’utilisation de ces outils offre un vernis d’objectivité qui masque non seulement un racisme et une xénophobie flagrants, mais aussi la forte pression politique exercée sur les agences d’immigration pour restreindre l’asile. « L‘IA permet aux agences fédérales de mener des contrôles d’immigration de manière profondément et de plus en plus opaque, ce qui complique encore davantage la tâche des personnes susceptibles d’être arrêtées ou accusées à tort. Nombre de ces outils ne sont connus du public que par le biais de documents juridiques et ne figurent pas dans l’inventaire d’IA du DHS. Mais même une fois connus, nous disposons de très peu d’informations sur leur étalonnage ou sur les données sur lesquelles ils sont basés, ce qui réduit encore davantage la capacité des individus à faire valoir leurs droits à une procédure régulière. Ces outils s’appuient également sur une surveillance invasive du public, allant du filtrage des publications sur les réseaux sociaux à l’utilisation de la reconnaissance faciale, de la surveillance aérienne et d’autres techniques de surveillance, à l’achat massif d’informations publiques auprès de courtiers en données ». Nous sommes à la fois confrontés à des systèmes coercitifs et opaques, foncièrement défaillants. Mais ces défaillances se déploient parce qu’elles donnent du pouvoir aux forces de l’ordre, leur permettant d’atteindre leurs objectifs d’expulsion et d’arrestation. Avec l’IA, le pouvoir devient l’objectif.

Les leviers pour renverser l’empire de l’IA et faire converger les luttes contre son monde

La dernière partie du rapport de l’AI Now Institute tente de déployer une autre vision de l’IA par des propositions, en dessinant une feuille de route pour l’action. “L’IA est une lutte de pouvoir et non un levier de progrès”, expliquent les auteurs qui invitent à “reprendre le contrôle de la trajectoire de l’IA”, en contestant son utilisation actuelle. Le rapport présente 5 leviers pour reprendre du pouvoir sur l’IA

Démontrer que l’IA agit contre les intérêts des individus et de la société

Le premier objectif, pour reprendre la main, consiste à mieux démontrer que l’industrie de l’IA agit contre les intérêts des citoyens ordinaires. Mais ce discours est encore peu partagé, notamment parce que le discours sur les risques porte surtout sur les biais techniques ou les risques existentiels, des enjeux déconnectés des réalités matérielles des individus. Pour l’AI Now Institute, “nous devons donner la priorité aux enjeux politiques ancrés dans le vécu des citoyens avec l’IA”, montrer les systèmes d’IA comme des infrastructures invisibles qui régissent les vies de chacun. En cela, la résistance au démantèlement des agences publiques initiée par les politiques du Doge a justement permis d’ouvrir un front de résistance. La résistance et l’indignation face aux coupes budgétaires et à l’accaparement des données a permis de montrer qu’améliorer l’efficacité des services n’était pas son objectif, que celui-ci a toujours été de démanteler les services gouvernementaux et centraliser le pouvoir. La dégradation des services sociaux et la privation des droits est un moyen de remobilisation à exploiter.

La construction des data centers pour l’IA est également un nouvel espace de mobilisation locale pour faire progresser la question de la justice environnementale, à l’image de celles que tentent de faire entendre la Citizen Action Coalition de l’Indiana ou la Memphis Community Against Pollution dans le Tennessee.

La question de l’augmentation des prix et de l’inflation, et le développements de prix et salaires algorithmiques est un autre levier de mobilisation, comme le montrait un rapport de l’AI Now Institute sur le sujet datant de février qui invitait à l’interdiction pure et simple de la surveillance individualisée des prix et des salaires. 

Faire progresser l’organisation des travailleurs 

Le second levier consiste à faire progresser l’organisation des travailleurs. Lorsque les travailleurs et leurs syndicats s’intéressent sérieusement à la manière dont l’IA transforme la nature du travail et s’engagent résolument par le biais de négociations collectives, de l’application des contrats, de campagnes et de plaidoyer politique, ils peuvent influencer la manière dont leurs employeurs développent et déploient ces technologies. Les campagnes syndicales visant à contester l’utilisation de l’IA générative à Hollywood, les mobilisations pour dénoncer la gestion algorithmique des employés des entrepôts de la logistique et des plateformes de covoiturage et de livraison ont joué un rôle essentiel dans la sensibilisation du public à l’impact de l’IA et des technologies de données sur le lieu de travail. La lutte pour limiter l’augmentation des cadences dans les entrepôts ou celles des chauffeurs menées par Gig Workers Rising, Los Deliversistas Unidos, Rideshare Drivers United, ou le SEIU, entre autres, a permis d’établir des protections, de lutter contre la précarité organisée par les plateformes… Pour cela, il faut à la fois que les organisations puissent analyser l’impact de l’IA sur les conditions de travail et sur les publics, pour permettre aux deux luttes de se rejoindre à l’image de ce qu’à accompli le syndicat des infirmières qui a montré que le déploiement de l’IA affaiblit le jugement clinique des infirmières et menace la sécurité des patients. Cette lutte a donné lieu à une « Déclaration des droits des infirmières et des patients », un ensemble de principes directeurs visant à garantir une application juste et sûre de l’IA dans les établissements de santé. Les infirmières ont stoppé le déploiement d’EPIC Acuity, un système qui sous-estimait l’état de santé des patients et le nombre d’infirmières nécessaires, et ont contraint l’entreprise qui déployait le système à créer un comité de surveillance pour sa mise en œuvre. 

Une autre tactique consiste à contester le déploiement d’IA austéritaires dans le secteur public à l’image du réseau syndicaliste fédéral, qui mène une campagne pour sauver les services fédéraux et met en lumière l’impact des coupes budgétaires du Doge. En Pennsylvanie, le SEIU a mis en place un conseil des travailleurs pour superviser le déploiement de solutions d’IA génératives dans les services publics. 

Une autre tactique consiste à mener des campagnes plus globales pour contester le pouvoir des grandes entreprises technologiques, comme la Coalition Athena qui demande le démantèlement d’Amazon, en reliant les questions de surveillance des travailleurs, le fait que la multinationale vende ses services à la police, les questions écologiques liées au déploiement des plateformes logistiques ainsi que l’impact des systèmes algorithmiques sur les petites entreprises et les prix que payent les consommateurs. 

Bref, l’enjeu est bien de relier les luttes entre elles, de relier les syndicats aux organisations de défense de la vie privée à celles œuvrant pour la justice raciale ou sociale, afin de mener des campagnes organisées sur ces enjeux. Mais également de l’étendre à l’ensemble de la chaîne de valeur et d’approvisionnement de l’IA, au-delà des questions américaines, même si pour l’instant “aucune tentative sérieuse d’organisation du secteur impacté par le déploiement de l’IA à grande échelle n’a été menée”. Des initiatives existent pourtant comme l’Amazon Employees for Climate Justice, l’African Content Moderators Union ou l’African Tech Workers Rising, le Data Worker’s Inquiry Project, le Tech Equity Collaborative ou l’Alphabet Workers Union (qui font campagne sur les différences de traitement entre les employés et les travailleurs contractuels). 

Nous avons désespérément besoin de projets de lutte plus ambitieux et mieux dotés en ressources, constate le rapport. Les personnes qui construisent et forment les systèmes d’IA – et qui, par conséquent, les connaissent intimement – ​​ont une opportunité particulière d’utiliser leur position de pouvoir pour demander des comptes aux entreprises technologiques sur la manière dont ces systèmes sont utilisés. “S’organiser et mener des actions collectives depuis ces postes aura un impact profond sur l’évolution de l’IA”.

“À l’instar du mouvement ouvrier du siècle dernier, le mouvement ouvrier d’aujourd’hui peut se battre pour un nouveau pacte social qui place l’IA et les technologies numériques au service de l’intérêt public et oblige le pouvoir irresponsable d’aujourd’hui à rendre des comptes.”

Confiance zéro envers les entreprises de l’IA !

Le troisième levier que défend l’AI Now Institute est plus radical encore puisqu’il propose d’adopter un programme politique “confiance zéro” envers l’IA. En 2023, L’AI Now, l’Electronic Privacy Information Center et d’Accountable Tech affirmaient déjà “qu’une confiance aveugle dans la bienveillance des entreprises technologiques n’était pas envisageable ». Pour établir ce programme, le rapport égraine 6 leviers à activer.

Tout d’abord, le rapport plaide pour “des règles audacieuses et claires qui restreignent les applications d’IA nuisibles”. C’est au public de déterminer si, dans quels contextes et comment, les systèmes d’IA seront utilisés. “Comparées aux cadres reposant sur des garanties basées sur les processus (comme les audits d’IA ou les régimes d’évaluation des risques) qui, dans la pratique, ont souvent eu tendance à renforcer les pouvoirs des leaders du secteur et à s’appuyer sur une solide capacité réglementaire pour une application efficace, ces règles claires présentent l’avantage d’être facilement administrables et de cibler les préjudices qui ne peuvent être ni évités ni réparés par de simples garanties”. Pour l’AI Now Institute, l’IA doit être interdite pour la reconnaissance des émotions, la notation sociale, la fixation des prix et des salaires, refuser des demandes d’indemnisation, remplacer les enseignants, générer des deepfakes. Et les données de surveillance des travailleurs ne doivent pas pouvoir pas être vendues à des fournisseurs tiers. L’enjeu premier est d’augmenter le spectre des interdictions. 

Ensuite, le rapport propose de réglementer tout le cycle de vie de l’IA. L’IA doit être réglementée tout au long de son cycle de développement, de la collecte des données au déploiement, en passant par le processus de formation, le perfectionnement et le développement des applications, comme le proposait l’Ada Lovelace Institute. Le rapport rappelle que si la transparence est au fondement d’une réglementation efficace, la résistante des entreprises est forte, tout le long des développements, des données d’entraînement utilisées, aux fonctionnement des applications. La transparence et l’explication devraient être proactives, suggère le rapport : les utilisateurs ne devraient pas avoir besoin de demander individuellement des informations sur les traitements dont ils sont l’objet. Notamment, le rapport insiste sur le besoin que “les développeurs documentent et rendent publiques leurs techniques d’atténuation des risques, et que le régulateur exige la divulgation de tout risque anticipé qu’ils ne sont pas en mesure d’atténuer, afin que cela soit transparent pour les autres acteurs de la chaîne d’approvisionnement”. Le rapport recommande également d’inscrire un « droit de dérogation » aux décisions et l’obligation d’intégrer des conseils d’usagers pour qu’ils aient leur mot à dire sur les développements et l’utilisation des systèmes. 

Le rapport rappelle également que la supervision des développements doit être indépendante. Ce n’est pas à l’industrie d’évaluer ce qu’elle fait. Le “red teaming” et les “models cards” ignorent les conflits d’intérêts en jeu et mobilisent des méthodologies finalement peu robustes (voir notre article). Autre levier encore, s’attaquer aux racines du pouvoir de ces entreprises et par exemple qu’elles suppriment les données acquises illégalement et les modèles entraînés sur ces données (certains chercheurs parlent d’effacement de modèles et de destruction algorithmique !) ; limiter la conservation des données pour le réentraînement ; limiter les partenariats entre les hyperscalers et les startups d’IA et le rachat d’entreprise pour limiter la constitution de monopoles

Le rapport propose également de construire une boîte à outils pour favoriser la concurrence. De nombreuses enquêtes pointent les limites des grandes entreprises de la tech à assurer le respect du droit à la concurrence, mais les poursuites peinent à s’appliquer et peinent à construire des changements législatifs pour renforcer le droit à la concurrence et limiter la construction de monopoles, alors que toute intervention sur le marché est toujours dénoncé par les entreprises de la tech comme relevant de mesures contre l’innovation. Le rapport plaide pour une plus grande séparation structurelle des activités (les entreprises du cloud ne doivent pas pouvoir participer au marché des modèles fondamentaux de l’IA par exemple, interdiction des représentations croisées dans les conseils d’administration des startups et des développeurs de modèles, etc.). Interdire aux fournisseurs de cloud d’exploiter les données qu’ils obtiennent de leurs clients en hébergeant des infrastructures pour développer des produits concurrents. 

Enfin, le rapport recommande une supervision rigoureuse du développement et de l’exploitation des centres de données, alors que les entreprises qui les développent se voient exonérées de charge et que leurs riverains en subissent des impacts disproportionnés (concurrence sur les ressources, augmentation des tarifs de l’électricité…). Les communautés touchées ont besoin de mécanismes de transparence et de protections environnementales solides. Les régulateurs devraient plafonner les subventions en fonction des protections concédées et des emplois créés. Initier des règles pour interdire de faire porter l’augmentation des tarifs sur les usagers.

Décloisonner !

Le cloisonnement des enjeux de l’IA est un autre problème qu’il faut lever. C’est le cas notamment de l’obsession à la sécurité nationale qui justifient à la fois des mesures de régulation et des programmes d’accélération et d’expansion du secteur et des infrastructures de l’IA. Mais pour décloisonner, il faut surtout venir perturber le processus de surveillance à l’œuvre et renforcer la vie privée comme un enjeu de justice économique. La montée de la surveillance pour renforcer l’automatisation “place les outils traditionnels de protection de la vie privée (tels que le consentement, les options de retrait, les finalités non autorisées et la minimisation des données) au cœur de la mise en place de conditions économiques plus justes”. La chercheuse Ifeoma Ajunwa soutient que les données des travailleurs devraient être considérées comme du « capital capturé » par les entreprises : leurs données sont  utilisées pour former des technologies qui finiront par les remplacer (ou créer les conditions pour réduire leurs salaires), ou vendues au plus offrant via un réseau croissant de courtiers en données, sans contrôle ni compensation. Des travailleurs ubérisés aux travailleurs du clic, l’exploitation des données nécessite de repositionner la protection de la vie privée des travailleurs au cœur du programme de justice économique pour limiter sa capture par l’IA. Les points de collecte, les points de surveillance, doivent être “la cible appropriée de la résistance”, car ils seront instrumentalisés contre les intérêts des travailleurs. Sur le plan réglementaire, cela pourrait impliquer de privilégier des règles de minimisation des données qui restreignent la collecte et l’utilisation des données, renforcer la confidentialité (par exemple en interdisant le partage de données sur les salariés avec des tiers), le droit à ne pas consentir, etc. Renforcer la minimisation, sécuriser les données gouvernementales sur les individus qui sont de haute qualité et particulièrement sensibles, est plus urgent que jamais. 

“Nous devons nous réapproprier l’agenda positif de l’innovation centrée sur le public, et l’IA ne devrait pas en être le centre”, concluent les auteurs. La trajectoire actuelle de l’IA, axée sur le marché, est préjudiciable au public alors que l’espace de solutions alternatives se réduit. Nous devons rejeter le paradigme d’une IA à grande échelle qui ne profitera qu’aux plus puissants.

L’IA publique demeure un espace fertile pour promouvoir le débat sur des trajectoires alternatives pour l’IA, structurellement plus alignées sur l’intérêt public, et garantir que tout financement public dans ce domaine soit conditionné à des objectifs d’intérêt général. Mais pour cela, encore faut-il que l’IA publique ne limite pas sa politique à l’achat de solutions privées, mais développe ses propres capacités d’IA, réinvestisse sa capacité d’expertise pour ne pas céder au solutionnisme de l’IA, favorise partout la discussion avec les usagers, cultive une communauté de pratique autour de l’innovation d’intérêt général qui façonnera l’émergence d’un espace alternatif par exemple en exigeant des méthodes d’implication des publics et aussi en élargissant l’intérêt de l’Etat à celui de l’intérêt collectif et pas seulement à ses intérêts propres (par exemple en conditionnant à la promotion des objectifs climatiques, au soutien syndical et citoyen…), ainsi qu’à redéfinir les conditions concrètes du financement public de l’IA, en veillant à ce que les investissements répondent aux besoins des communautés plutôt qu’aux intérêts des entreprises.   

Changer l’agenda : pour une IA publique !

Enfin, le rapport conclut en affirmant que l’innovation devrait être centrée sur les besoins des publics et que l’IA ne devrait pas en être le centre. Le développement de l’IA devrait être guidé par des impératifs non marchands et les capitaux publics et philanthropiques devraient contribuer à la création d’un écosystème d’innovation extérieur à l’industrie, comme l’ont réclamé Public AI Network dans un rapport, l’Ada Lovelace Institute, dans un autre, Lawrence Lessig ou encore Bruce Schneier et Nathan Sanders ou encore Ganesh Sitaraman et Tejas N. Narechania…  qui parlent d’IA publique plus que d’IA souveraine, pour orienter les investissement non pas tant vers des questions de sécurité nationale et de compétitivité, mais vers des enjeux de justice sociale. 

Ces discours confirment que la trajectoire de l’IA, axée sur le marché, est préjudiciable au public. Si les propositions alternatives ne manquent pas, elles ne parviennent pas à relever le défi de la concentration du pouvoir au profit des grandes entreprises. « Rejeter le paradigme actuel de l’IA à grande échelle est nécessaire pour lutter contre les asymétries d’information et de pouvoir inhérentes à l’IA. C’est la partie cachée qu’il faut exprimer haut et fort. C’est la réalité à laquelle nous devons faire face si nous voulons rassembler la volonté et la créativité nécessaires pour façonner la situation différemment ». Un rapport du National AI Research Resource (NAIRR) américain de 2021, d’une commission indépendante présidée par l’ancien PDG de Google, Eric Schmidt, et composée de dirigeants de nombreuses grandes entreprises technologiques, avait parfaitement formulé le risque : « la consolidation du secteur de l’IA menace la compétitivité technologique des États-Unis. » Et la commission proposait de créer des ressources publiques pour l’IA. 

« L’IA publique demeure un espace fertile pour promouvoir le débat sur des trajectoires alternatives pour l’IA, structurellement plus alignées sur l’intérêt général, et garantir que tout financement public dans ce domaine soit conditionné à des objectifs d’intérêt général ». Un projet de loi californien a récemment relancé une proposition de cluster informatique public, hébergé au sein du système de l’Université de Californie, appelé CalCompute. L’État de New York a lancé une initiative appelée Empire AI visant à construire une infrastructure de cloud public dans sept institutions de recherche de l’État, rassemblant plus de 400 millions de dollars de fonds publics et privés. Ces deux initiatives créent des espaces de plaidoyer importants pour garantir que leurs ressources répondent aux besoins des communautés et ne servent pas à enrichir davantage les ressources des géants de la technologie.

Et le rapport de se conclure en appelant à défendre l’IA publique, en soutenant les universités, en investissant dans ces infrastructures d’IA publique et en veillant que les groupes défavorisés disposent d’une autorité dans ces projets. Nous devons cultiver une communauté de pratique autour de l’innovation d’intérêt général. 

***

Le rapport de l’AI Now Institute a la grande force de nous rappeler que les luttes contre l’IA existent et qu’elles ne sont pas que des luttes de collectifs technocritiques, mais qu’elles s’incarnent déjà dans des projets politiques, qui peinent à s’interelier et à se structurer. Des luttes qui sont souvent invisibilisées, tant la parole est toute entière donnée aux promoteurs de l’IA. Le rapport est extrêmement riche et rassemble une documentation à nulle autre pareille. 

« L’IA ne nous promet ni de nous libérer du cycle incessant de guerres, des pandémies et des crises environnementales et financières qui caractérisent notre présent », conclut le rapport  L’IA ne crée rien de tout cela, ne créé rien de ce que nous avons besoin. “Lier notre avenir commun à l’IA rend cet avenir plus difficile à réaliser, car cela nous enferme dans une voie résolument sombre, nous privant non seulement de la capacité de choisir quoi construire et comment le construire, mais nous privant également de la joie que nous pourrions éprouver à construire un avenir différent”. L’IA comme seule perspective d’avenir “nous éloigne encore davantage d’une vie digne, où nous aurions l’autonomie de prendre nos propres décisions et où des structures démocratiquement responsables répartiraient le pouvoir et les infrastructures technologiques de manière robuste, responsable et protégée des chocs systémiques”. L’IA ne fait que consolider et amplifier les asymétries de pouvoir existantes. “Elle naturalise l’inégalité et le mérite comme une fatalité, ​tout en rendant les schémas et jugements sous-jacents qui les façonnent impénétrables pour ceux qui sont affectés par les jugements de l’IA”.

Pourtant, une autre IA est possible, estiment les chercheurs.ses de l’AI Now Institute. Nous ne pouvons pas lutter contre l’oligarchie technologique sans rejeter la trajectoire actuelle de l’industrie autour de l’IA à grande échelle. Nous ne devons pas oublier que l’opinion publique s’oppose résolument au pouvoir bien établi des entreprises technologiques. Certes, le secteur technologique dispose de ressources plus importantes que jamais et le contexte politique est plus sombre que jamais, concèdent les chercheurs de l’AI Now Institute. Cela ne les empêche pas de faire des propositions, comme d’adopter un programme politique de « confiance zéro » pour l’IA. Adopter un programme politique fondé sur des règles claires qui restreignent les utilisations les plus néfastes de l’IA, encadrent son cycle de vie de bout en bout et garantissent que l’industrie qui crée et exploite actuellement l’IA ne soit pas laissée à elle-même pour s’autoréguler et s’autoévaluer. Repenser les leviers traditionnels de la confidentialité des données comme outils clés dans la lutte contre l’automatisation et la lutte contre le pouvoir de marché.

Revendiquer un programme positif d’innovation centrée sur le public, sans IA au centre. 

« La trajectoire actuelle de l’IA place le public sous la coupe d’oligarques technologiques irresponsables. Mais leur succès n’est pas inéluctable. En nous libérant de l’idée que l’IA à grande échelle est inévitable, nous pouvons retrouver l’espace nécessaire à une véritable innovation et promouvoir des voies alternatives stimulantes et novatrices qui exploitent la technologie pour façonner un monde au service du public et gouverné par notre volonté collective ».

La trajectoire actuelle de l’IA vers sa suprématie ne nous mènera pas au monde que nous voulons. Sa suprématie n’est pourtant pas encore là. “Avec l’adoption de la vision actuelle de l’IA, nous perdons un avenir où l’IA favoriserait des emplois stables, dignes et valorisants. Nous perdons un avenir où l’IA favoriserait des salaires justes et décents, au lieu de les déprécier ; où l’IA garantirait aux travailleurs le contrôle de l’impact des nouvelles technologies sur leur carrière, au lieu de saper leur expertise et leur connaissance de leur propre travail ; où nous disposons de politiques fortes pour soutenir les travailleurs si et quand les nouvelles technologies automatisent les fonctions existantes – y compris des lois élargissant le filet de sécurité sociale – au lieu de promoteurs de l’IA qui se vantent auprès des actionnaires des économies réalisées grâce à l’automatisation ; où des prestations sociales et des politiques de congés solides garantissent le bien-être à long terme des employés, au lieu que l’IA soit utilisée pour surveiller et exploiter les travailleurs à tout va ; où l’IA contribue à protéger les employés des risques pour la santé et la sécurité au travail, au lieu de perpétuer des conditions de travail dangereuses et de féliciter les employeurs qui exploitent les failles du marché du travail pour se soustraire à leurs responsabilités ; et où l’IA favorise des liens significatifs par le travail, au lieu de favoriser des cultures de peur et d’aliénation.”

Pour l’AI Now Institute, l’enjeu est d’aller vers une prospérité partagée, et ce n’est pas la direction que prennent les empires de l’IA. La prolifération de toute nouvelle technologie a le potentiel d’accroître les opportunités économiques et de conduire à une prospérité partagée généralisée. Mais cette prospérité partagée est incompatible avec la trajectoire actuelle de l’IA, qui vise à maximiser le profit des actionnaires. “Le mythe insidieux selon lequel l’IA mènera à la « productivité » pour tous, alors qu’il s’agit en réalité de la productivité d’un nombre restreint d’entreprises, nous pousse encore plus loin sur la voie du profit actionnarial comme unique objectif économique. Même les politiques gouvernementales bien intentionnées, conçues pour stimuler le secteur de l’IA, volent les poches des travailleurs. Par exemple, les incitations gouvernementales destinées à revitaliser l’industrie de la fabrication de puces électroniques ont été contrecarrées par des dispositions de rachat d’actions par les entreprises, envoyant des millions de dollars aux entreprises, et non aux travailleurs ou à la création d’emplois. Et malgré quelques initiatives significatives pour enquêter sur le secteur de l’IA sous l’administration Biden, les entreprises restent largement incontrôlées, ce qui signifie que les nouveaux entrants ne peuvent pas contester ces pratiques.”

“Cela implique de démanteler les grandes entreprises, de restructurer la structure de financement financée par le capital-risque afin que davantage d’entreprises puissent prospérer, d’investir dans les biens publics pour garantir que les ressources technologiques ne dépendent pas des grandes entreprises privées, et d’accroître les investissements institutionnels pour intégrer une plus grande diversité de personnes – et donc d’idées – au sein de la main-d’œuvre technologique.”

“Nous méritons un avenir technologique qui soutienne des valeurs et des institutions démocratiques fortes.” Nous devons de toute urgence restaurer les structures institutionnelles qui protègent les intérêts du public contre l’oligarchie. Cela nécessitera de s’attaquer au pouvoir technologique sur plusieurs fronts, et notamment par la mise en place de mesures de responsabilisation des entreprises pour contrôler les oligarques de la tech. Nous ne pouvons les laisser s’accaparer l’avenir. 

Sur ce point, comme sur les autres, nous sommes d’accord.

Hubert Guillaud

  • ✇404 Media
  • Polymarket Gamblers Go to War Over Whether Zelenskyy Wore a Suit
    Polymarket, an online betting marketplace that bills itself as the future of news, can’t decide whether or not Ukrainian president Volodomyr Zelenskyy wore a suit during a recent appearance in Europe. The gambling site is set to make a final judgement about the question in a few hours and more than $160 million in crypto is riding on it.Polymarket is a gambling website where users predict the outcome of binary events. It gained prominence in the runup to the 2024 election, signed an exclusivi
     

Polymarket Gamblers Go to War Over Whether Zelenskyy Wore a Suit

7 juillet 2025 à 15:28
Polymarket Gamblers Go to War Over Whether Zelenskyy Wore a Suit

Polymarket, an online betting marketplace that bills itself as the future of news, can’t decide whether or not Ukrainian president Volodomyr Zelenskyy wore a suit during a recent appearance in Europe. The gambling site is set to make a final judgement about the question in a few hours and more than $160 million in crypto is riding on it.

Polymarket is a gambling website where users predict the outcome of binary events. It gained prominence in the runup to the 2024 election, signed an exclusivity deal with X in June, and sees itself not just as an online betting parlor, but as an arbiter of truth. Its founder, Shayne Coplan, thinks that the future of media belongs to a website made for degenerate gamblers to make silly bets.

And yet this arbiter of truth had trouble figuring out if Zelenskyy wore a suit at the end of June during a NATO summit. The bet, started on May 22, is simple: “Will Zelenskyy wear a suit before July?” The answer, it turns out, is pretty hard. When Zelenskyy showed up at a NATO summit wearing a tailored jacket and a button up shirt, a stark contrast to his more casual military style garb, a community-run Polymarket account posted, “President Zelenskyy in a suit last night.” 

  • ✇404 Media
  • The Open-Source Software Saving the Internet From AI Bot Scrapers
    For someone who says she is fighting AI bot scrapers just in her free time, Xe Iaso seems to be putting up an impressive fight. Since she launched it in January, Anubis, a “program is designed to help protect the small internet from the endless storm of requests that flood in from AI companies,” has been downloaded nearly 200,000 times, and is being used by notable organizations including GNOME, the popular open-source desktop environment for Linux, FFmpeg, the open-source software project for h
     

The Open-Source Software Saving the Internet From AI Bot Scrapers

7 juillet 2025 à 09:16
The Open-Source Software Saving the Internet From AI Bot Scrapers

For someone who says she is fighting AI bot scrapers just in her free time, Xe Iaso seems to be putting up an impressive fight. Since she launched it in January, Anubis, a “program is designed to help protect the small internet from the endless storm of requests that flood in from AI companies,” has been downloaded nearly 200,000 times, and is being used by notable organizations including GNOME, the popular open-source desktop environment for Linux, FFmpeg, the open-source software project for handling video and other media, and UNESCO, the United Nations organization for educations, science, and culture. 

Iaso decided to develop Anubis after discovering that her own Git server was struggling with AI scrapers, bots that crawl the web hoovering up anything that can be used for the training data that power AI models. Like many libraries, archives, and other small organizations, Iaso discovered her Git server was getting slammed only when it stopped working.  

  • ✇Dans les algorithmes
  • Pour lutter contre la désinformation, il faut reconstruire du social
    L’Institut Nicod publie un court et très stimulant rapport sur la désinformation signé Grégoire Darcy. Non seulement celui-ci débogue la simplicité des réponses cognitives que les politiques publiques ont tendance à proposer, mais surtout, repolitise la question.  Le rapport rappelle que la désinformation n’est pas seulement un problème d’irrationnalité et de crédulité. Il invite à sortir de l’approche réactive qui se concentre sur les symptômes et qui se focalise bien trop sur les modalités
     

Pour lutter contre la désinformation, il faut reconstruire du social

7 juillet 2025 à 01:00

L’Institut Nicod publie un court et très stimulant rapport sur la désinformation signé Grégoire Darcy. Non seulement celui-ci débogue la simplicité des réponses cognitives que les politiques publiques ont tendance à proposer, mais surtout, repolitise la question. 

Le rapport rappelle que la désinformation n’est pas seulement un problème d’irrationnalité et de crédulité. Il invite à sortir de l’approche réactive qui se concentre sur les symptômes et qui se focalise bien trop sur les modalités de diffusion oubliant les mécanismes affectifs et sociaux qui expliquent l’adhésion aux récits trompeurs. La lutte contre la désinformation repose sur une vision simpliste de la psychologie humaine : « la désinformation répond à des besoins sociaux, émotionnels et identitaires plus qu’à de simples déficits de rationalité. Ainsi, corriger les erreurs factuelles ne suffit pas : il faut s’attaquer aux conditions qui rendent ces récits socialement fonctionnels. » La désinformation n’est que le symptôme de la dégradation globale de l’écosystème informationnel. « Les vulnérabilités face à la désinformation ne tiennent pas qu’aux dispositions individuelles, mais s’ancrent dans des environnements sociaux, économiques et médiatiques spécifiques : isolement social, précarité, homogamie idéologique et défiance institutionnelle sont des facteurs clés expliquant l’adhésion, bien au-delà des seuls algorithmes ou biais cognitifs ».

“Tant que les politiques publiques se contenteront de réponses réactives, centrées sur les symptômes visibles et ignorantes des dynamiques cognitives, sociales et structurelles à l’œuvre, elles risquent surtout d’aggraver ce qu’elles prétendent corriger. En cause : un modèle implicite, souvent naïf, de la psychologie humaine – un schéma linéaire et individualisant, qui réduit l’adhésion aux contenus trompeurs à un simple déficit d’information ou de rationalité. Ce cadre conduit à des politiques fragmentées, peu efficaces, parfois même contre-productive.” 

Les réponses les plus efficientes à la désinformation passent par une transformation structurelle de l’écosystème informationnel, que seule l’action publique peut permettre, en orchestrant à la fois la régulation algorithmique et le renforcement des médias fiables. La réduction des vulnérabilités sociales, économiques et institutionnelles constitue l’approche la plus structurante pour lutter contre la désinformation, en s’attaquant aux facteurs qui nourrissent la réceptivité aux contenus trompeurs – précarité, marginalisation, polarisation et défiance envers les institutions. Parmi les mesures que pointe le rapport, celui-ci invite à une régulation forte des réseaux sociaux permettant de « restituer la maîtrise du fil par une transparence algorithmique accrue et une possibilité de maîtriser » les contenus auxquels les gens accèdent : « rendre visibles les critères de recommandation et proposer par défaut un fil chronologique permettrait de réduire les manipulations attentionnelles sans recourir à la censure ». Le rapport recommande également « d’assurer un financement stable pour garantir l’indépendance des médias et du service public d’information ». Il recommande également de renforcer la protection sociale et les politiques sociales pour renforcer la stabilité propice à l’analyse critique. D’investir dans le développement d’espace de sociabilité et de favoriser une circulation apaisée de l’information en renforçant l’intégrité publique. 

Un rapport stimulant, qui prend à rebours nos présupposés et qui nous dit que pour lutter contre la désinformation, il faut lutter pour rétablir une société juste.

  • ✇404 Media
  • Podcast: The Life Changing Power of Lifting
    For this week’s podcast, I’m talking to our friend Casey Johnston, a tech journalist turned fitness journalist turned independent journalist. Casey studied physics, which led her to tech journalism; she did some of my favorite coverage of Internet culture as well as Apple’s horrendous butterfly laptop keyboards. We worked together at VICE, where Casey was an editor and where she wrote Ask a Swole Woman, an advice column about weightlifting. After she left VICE, Casey founded She’s a Beast, an
     

Podcast: The Life Changing Power of Lifting

3 juillet 2025 à 06:00
Podcast: The Life Changing Power of Lifting

For this week’s podcast, I’m talking to our friend Casey Johnston, a tech journalist turned fitness journalist turned independent journalist. Casey studied physics, which led her to tech journalism; she did some of my favorite coverage of Internet culture as well as Apple’s horrendous butterfly laptop keyboards. We worked together at VICE, where Casey was an editor and where she wrote Ask a Swole Woman, an advice column about weightlifting. After she left VICE, Casey founded She’s a Beast, an independent site about weightlifting, but also about the science of diet culture, fitness influencers on the internet, the intersections of all those things, etc. 

She just wrote A Physical Education: How I Escaped Diet Culture and Gained the Power of Lifting, a really great reported memoir about how our culture and the media often discourages people from lifting, and how this type of exercise can be really beneficial to your brain and your body. I found the book really inspiring and actually started lifting right after I read it. In this interview we talk about her book, about journalism, about independent media, and how doing things like lifting weights and touching grass helps us navigate the world.

Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a paid subscriber, check your inbox for an email from our podcast host Transistor for a link to the subscribers-only version! You can also add that subscribers feed to your podcast app of choice and never miss an episode that way. The email should also contain the subscribers-only unlisted YouTube link for the extended video version too. It will also be in the show notes in your podcast player.

  • ✇Dans les algorithmes
  • La santé au prisme de son abandon
    Dans AOC, le philosophe Alexandre Monnin, auteur de Politiser le renoncement (Divergences, 2023) explique que “derrière les discours d’efficience, d’autonomie et de prévention, un glissement insidieux s’opère : celui d’une médecine qui renonce à soigner”. Le soin est en train de devenir conditionnel, réservé aux existences jugées “optimisables”. La stratégie de non-soin, n’est pas que la conséquence des restrictions budgétaires ou de la désorganisation du secteur, mais une orientation active, un
     

La santé au prisme de son abandon

3 juillet 2025 à 00:55

Dans AOC, le philosophe Alexandre Monnin, auteur de Politiser le renoncement (Divergences, 2023) explique que “derrière les discours d’efficience, d’autonomie et de prévention, un glissement insidieux s’opère : celui d’une médecine qui renonce à soigner”. Le soin est en train de devenir conditionnel, réservé aux existences jugées “optimisables”. La stratégie de non-soin, n’est pas que la conséquence des restrictions budgétaires ou de la désorganisation du secteur, mais une orientation active, un projet politique. Comme c’est le cas au travers du programme américain MAHA (Make America Healthy Again), dont l’ambien n’est plus de soigner, mais d’éviter les coûts liés au soin, ou la loi sur le droit à mourir récemment adoptée en France, dénoncée par les collectifs antivalidistes comme une manière d’acter l’impossibilité de vivre avec certains handicaps ou maladies chroniques. “Ce tournant ne se donne pas toujours pour ce qu’il est. Il s’abrite derrière les mots d’efficacité, d’autonomie, de prévention, voire de soutenabilité. Il s’appuie sur des cadres comme le paradigme One Health, censé penser la santé de manière systémique à l’échelle des écosystèmes mais qui, en pratique, contribue à diluer les responsabilités et à rendre invisibles les enjeux de justice sociale.” Nous entrons dans une médicalisation sans soins, où l’analyse de santé se détache de toute thérapeutique.

Pour Derek Beres de Conspirituality, nous entrons dans une ère de soft eugenics”, d’eugénisme doux. Le self-care propose désormais à chacun de mesurer sa santé pour en reprendre le contrôle, dans une forme de “diagnostics sans soins”, qui converge avec les vues antivax de Robert Kennedy Jr, le ministre de la Santé américain, critiquant à la fois la surmédicalisation et la montée des maladies chroniques renvoyées à des comportements individuels. En mettant l’accent sur la prévention et la modification des modes de vies, cet abandon de la santé renvoie les citoyens vers leurs responsabilités et la santé publique vers des solutions privées, en laissant sur le carreau les populations vulnérables. Cette médecine du non-soin s’appuie massivement sur des dispositifs technologiques sophistiqués proches du quantified self, “vidée de toute relation clinique”. “Ces technologies alimentent des systèmes d’optimisation où l’important n’est plus la guérison, mais la conformité aux normes biologiques ou comportementales. Dans ce contexte, le patient devient un profil de risque, non plus un sujet à accompagner. La plateformisation du soin réorganise en profondeur les régimes d’accès à la santé. La médecine n’est alors plus un service public mais une logistique de gestion différenciée des existences.”

C’est le cas du paradigme One Health, qui vise à remplacer le soin par une idéalisation holistique de la santé, comme un état d’équilibre à maintenir, où l’immunité naturelle affaiblit les distinctions entre pathogène et environnement et favorise une démission institutionnelle. “Face aux dégradations écologiques, le réflexe n’est plus de renforcer les capacités collectives de soin. Il s’agit désormais de retrouver une forme de pureté corporelle ou environnementale perdue. Cette quête se traduit par l’apologie du jeûne, du contact avec les microbes, de la « vitalité » naturelle – et la dénonciation des traitements, des masques, des vaccins comme autant d’artefacts « toxiques ». Elle entretient une confusion entre médecine industrielle et médecine publique, et reformule le soin comme une purification individuelle. Là encore, le paradigme du non-soin prospère non pas en contradiction avec l’écologie, mais bien davantage au nom d’une écologie mal pensée, orientée vers le refus de l’artifice plutôt que vers l’organisation solidaire de la soutenabilité.” “L’appel à « ne pas tomber malade » devient un substitut direct au droit au soin – voire une norme visant la purification des plus méritants dans un monde saturé de toxicités (et de modernité).”

“Dans ce monde du non-soin, l’abandon n’est ni un effet secondaire ni une faute mais un principe actif de gestion.” Les populations vulnérables sont exclues de la prise en charge. Sous forme de scores de risques, le tri sanitaire technicisé s’infiltre partout, pour distinguer les populations et mettre de côté ceux qui ne peuvent être soignés. “La santé publique cesse d’être pensée comme un bien commun, et devient une performance individuelle, mesurée, scorée, marchandée. La médecine elle-même, soumise à l’austérité, finit par abandonner ses missions fondamentales : observer, diagnostiquer, soigner. Elle se contente de prévenir – et encore, seulement pour ceux qu’on juge capables – et/ou suffisamment méritants.” Pour Monnin, cet accent mis sur la prévention pourrait être louable si elle ne se retournait pas contre les malades : “Ce n’est plus la santé publique qui se renforce mais une responsabilité individualisée du « bien se porter » qui légitime l’abandon de celles et ceux qui ne peuvent s’y conformer. La prévention devient une rhétorique de la culpabilité, où le soin est indexé sur la conformité à un mode de vie puissamment normé”.

Pour le philosophe, le risque est que le soin devienne une option, un privilège.

Le problème est que ces nouvelles politiques avancent sous le masque de l’innovation et de la prévention, alors qu’elles ne parlent que de responsabilité individuelle, au risque de faire advenir un monde sans soin qui refuse d’intervenir sur les milieux de vies, qui refuse les infrastructures collectives, qui renvoie chacun à l’auto-surveillance “sans jamais reconstruire les conditions collectives du soin ni reconnaître l’inégale capacité des individus à le faire”. Un monde où ”la surveillance remplace l’attention, la donnée remplace la relation, le test remplace le soin”. Derrière le tri, se profile “une santé sans soin, une médecine sans clinique – une écologie sans solidarité”.

“L’État ne disparaît pas : il prescrit, organise, finance, externalise. Il se fait plateforme, courtier de services, émetteur d’appels à projets. En matière de santé, cela signifie le financement de dispositifs de prévention algorithmique, l’encouragement de solutions « innovantes » portées par des start-ups, ou encore le remboursement indirect de produits encore non éprouvés. Ce nouveau régime n’est pas une absence de soin, c’est une délégation programmée du soin à des acteurs dont l’objectif premier n’est pas le soin mais la rentabilité. L’État ne s’efface pas en totalité : il administre la privatisation du soin.”

  • ✇Dans les algorithmes
  • IA et éducation (2/2) : du dilemme moral au malaise social
    Suite de notre dossier sur IA et éducation (voir la première partie). La bataille éducative est-elle perdue ? Une grande enquête de 404 media montre qu’à l’arrivée de ChatGPT, les écoles publiques américaines étaient totalement démunies face à l’adoption généralisée de ChatGPT par les élèves. Le problème est d’ailleurs loin d’être résolu. Le New York Mag a récemment publié un article qui se désole de la triche généralisée qu’ont introduit les IA génératives à l’école. De partout, les élèv
     

IA et éducation (2/2) : du dilemme moral au malaise social

1 juillet 2025 à 00:41

Suite de notre dossier sur IA et éducation (voir la première partie).

La bataille éducative est-elle perdue ?

Une grande enquête de 404 media montre qu’à l’arrivée de ChatGPT, les écoles publiques américaines étaient totalement démunies face à l’adoption généralisée de ChatGPT par les élèves. Le problème est d’ailleurs loin d’être résolu. Le New York Mag a récemment publié un article qui se désole de la triche généralisée qu’ont introduit les IA génératives à l’école. De partout, les élèves utilisent les chatbots pour prendre des notes pendant les cours, pour concevoir des tests, résumer des livres ou des articles, planifier et rédiger leurs essais, résoudre les exercices qui leurs sont demandés. Le plafond de la triche a été pulvérisé, explique un étudiant. “Un nombre considérable d’étudiants sortiront diplômés de l’université et entreront sur le marché du travail en étant essentiellement analphabètes”, se désole un professeur qui constate le court-circuitage du processus même d’apprentissage. La triche semblait pourtant déjà avoir atteint son apogée, avant l’arrivée de ChatGPT, notamment avec les plateformes d’aides au devoir en ligne comme Chegg et Course Hero. “Pour 15,95 $ par mois, Chegg promettait des réponses à toutes les questions de devoirs en seulement 30 minutes, 24h/24 et 7j/7, grâce aux 150 000 experts diplômés de l’enseignement supérieur qu’elle employait, principalement en Inde”

Chaque école a proposé sa politique face à ces nouveaux outils, certains prônant l’interdiction, d’autres non. Depuis, les politiques se sont plus souvent assouplies, qu’endurcies. Nombre de profs autorisent l’IA, à condition de la citer, ou ne l’autorisent que pour aide conceptuelle et en demandant aux élèves de détailler la manière dont ils l’ont utilisé. Mais cela ne dessine pas nécessairement de limites claires à leurs usages. L’article souligne que si les professeurs se croient doués pour détecter les écrits générés par l’IA, des études ont démontré qu’ils ne le sont pas. L’une d’elles, publiée en juin 2024, utilisait de faux profils d’étudiants pour glisser des travaux entièrement générés par l’IA dans les piles de correction des professeurs d’une université britannique. Les professeurs n’ont pas signalé 97 % des essais génératifs. En fait, souligne l’article, les professeurs ont plutôt abandonné l’idée de pouvoir détecter le fait que les devoirs soient rédigés par des IA. “De nombreux enseignants semblent désormais désespérés”. “Ce n’est pas ce pour quoi nous nous sommes engagés”, explique l’un d’entre eux. La prise de contrôle de l’enseignement par l’IA tient d’une crise existentielle de l’éducation. Désormais, les élèves ne tentent même plus de se battre contre eux-mêmes. Ils se replient sur la facilité. “Toute tentative de responsabilisation reste vaine”, constatent les professeurs. 

L’IA a mis à jour les défaillances du système éducatif. Bien sûr, l’idéal de l’université et de l’école comme lieu de développement intellectuel, où les étudiants abordent des idées profondes a disparu depuis longtemps. La perspective que les IA des professeurs évaluent désormais les travaux produits par les IA des élèves, finit de réduire l’absurdité de la situation, en laissant chacun sans plus rien à apprendre. Plusieurs études (comme celle de chercheurs de Microsoft) ont établi un lien entre l’utilisation de l’IA et une détérioration de l’esprit critique. Pour le psychologue, Robert Sternberg, l’IA générative compromet déjà la créativité et l’intelligence. “La bataille est perdue”, se désole un autre professeur

Reste à savoir si l’usage “raisonnable” de l’IA est possible. Dans une longue enquête pour le New Yorker, le journaliste Hua Hsu constate que tous les étudiants qu’il a interrogé pour comprendre leur usage de l’IA ont commencé par l’utiliser pour se donner des idées, en promettant de veiller à un usage responsable et ont très vite basculé vers des usages peu modérés, au détriment de leur réflexion. L’utilisation judicieuse de l’IA ne tient pas longtemps. Dans un rapport sur l’usage de Claude par des étudiants, Anthropic a montré que la moitié des interactions des étudiants avec son outil serait extractive, c’est-à-dire servent à produire des contenus. 404 media est allé discuter avec les participants de groupes de soutien en ligne de gens qui se déclarent comme “dépendants à l’IA”. Rien n’est plus simple que de devenir accro à un chatbot, confient des utilisateurs de tout âge. OpenAI en est conscient, comme le pointait une étude du MIT sur les utilisateurs les plus assidus, sans proposer pourtant de remèdes.

Comment apprendre aux enfants à faire des choses difficiles ? Le journaliste Clay Shirky, devenu responsable de l’IA en éducation à la New York University, dans le Chronicle of Higher Education, s’interroge : l’IA améliore-t-elle l’éducation ou la remplace-t-elle ? “Chaque année, environ 15 millions d’étudiants de premier cycle aux États-Unis produisent des travaux et des examens de plusieurs milliards de mots. Si le résultat d’un cours est constitué de travaux d’étudiants (travaux, examens, projets de recherche, etc.), le produit de ce cours est l’expérience étudiante. Un devoir n’a de valeur que ”pour stimuler l’effort et la réflexion de l’élève”. “L’utilité des devoirs écrits repose sur deux hypothèses : la première est que pour écrire sur un sujet, l’élève doit comprendre le sujet et organiser ses pensées. La seconde est que noter les écrits d’un élève revient à évaluer l’effort et la réflexion qui y ont été consacrés”. Avec l’IA générative, la logique de cette proposition, qui semblait pourtant à jamais inébranlable, s’est complètement effondrée

Pour Shirky, il ne fait pas de doute que l’IA générative peut être utile à l’apprentissage. “Ces outils sont efficaces pour expliquer des concepts complexes, proposer des quiz pratiques, des guides d’étude, etc. Les étudiants peuvent rédiger un devoir et demander des commentaires, voir à quoi ressemble une réécriture à différents niveaux de lecture, ou encore demander un résumé pour vérifier la clart锓Mais le fait que l’IA puisse aider les étudiants à apprendre ne garantit pas qu’elle le fera. Pour le grand théoricien de l’éducation, Herbert Simon, “l’enseignant ne peut faire progresser l’apprentissage qu’en incitant l’étudiant à apprendre”. “Face à l’IA générative dans nos salles de classe, la réponse évidente est d’inciter les étudiants à adopter les utilisations utiles de l’IA tout en les persuadant d’éviter les utilisations néfastes. Notre problème est que nous ne savons pas comment y parvenir”, souligne pertinemment Shirky. Pour lui aussi, aujourd’hui, les professeurs sont en passe d’abandonner. Mettre l’accent sur le lien entre effort et apprentissage ne fonctionne pas, se désole-t-il. Les étudiants eux aussi sont déboussolés et finissent par se demander où l’utilisation de l’IA les mène. Shirky fait son mea culpa. L’utilisation engagée de l’IA conduit à son utilisation paresseuse. Nous ne savons pas composer avec les difficultés. Mais c’était déjà le cas avant ChatGPT. Les étudiants déclarent régulièrement apprendre davantage grâce à des cours magistraux bien présentés qu’avec un apprentissage plus actif, alors que de nombreuses études démontrent l’inverse. “Un outil qui améliore le rendement mais dégrade l’expérience est un mauvais compromis”. 

C’est le sens même de l’éducation qui est en train d’être perdu. Le New York Times revenait récemment sur le fait que certaines écoles interdisent aux élèves d’utiliser ces outils, alors que les professeurs, eux, les surutilisent. Selon une étude auprès de 1800 enseignants de l’enseignement supérieur, 18 % déclaraient utiliser fréquemment ces outils pour faire leur cours, l’année dernière – un chiffre qui aurait doublé depuis. Les étudiants ne lisent plus ce qu’ils écrivent et les professeurs non plus. Si les profs sont prompts à critiquer l’usage de l’IA par leurs élèves, nombre d’entre eux l’apprécient pour eux-mêmes, remarque un autre article du New York Times. A PhotoMath ou Google Lens qui viennent aider les élèves, répondent MagicSchool et Brisk Teaching qui proposent déjà des produits d’IA qui fournissent un retour instantané sur les écrits des élèves. L’Etat du Texas a signé un contrat de 5 ans avec l’entreprise Cambium Assessment pour fournir aux professeurs un outil de notation automatisée des écrits des élèves. 

Pour Jason Koebler de 404 media : “la société dans son ensemble n’a pas très bien résisté à l’IA générative, car les grandes entreprises technologiques s’obstinent à nous l’imposer. Il est donc très difficile pour un système scolaire public sous-financé de contrôler son utilisation”. Pourtant, peu après le lancement public de ChatGPT, certains districts scolaires locaux et d’État ont fait appel à des consultants pro-IA pour produire des formations et des présentations “encourageant largement les enseignants à utiliser l’IA générative en classe”, mais “aucun n’anticipait des situations aussi extrêmes que celles décrites dans l’article du New York Mag, ni aussi problématiques que celles que j’ai entendues de mes amis enseignants, qui affirment que certains élèves désormais sont totalement dépendants de ChatGPT”. Les documents rassemblés par 404media montrent surtout que les services d’éducation américains ont tardé à réagir et à proposer des perspectives aux enseignants sur le terrain. 

Dans un autre article de 404 media, Koebler a demandé à des professeurs américains d’expliquer ce que l’IA a changé à leur travail. Les innombrables témoignages recueillis montrent que les professeurs ne sont pas restés les bras ballants, même s’ils se sentent très dépourvus face à l’intrusion d’une technologie qu’ils n’ont pas voulu. Tous expliquent qu’ils passent des heures à corriger des devoirs que les élèves mettent quelques secondes à produire. Tous dressent un constat similaire fait d’incohérences, de confusions, de démoralisations, entre préoccupations et exaspérations. Quelles limites mettre en place ? Comment s’assurer qu’elles soient respectées ? “Je ne veux pas que les étudiants qui n’utilisent pas de LLM soient désavantagés. Et je ne veux pas donner de bonnes notes à des étudiants qui ne font pratiquement rien”, témoigne un prof. Beaucoup ont désormais recours à l’écriture en classe, au papier. Quelques-uns disent qu’ils sont passés de la curiosité au rejet catégorique de ces outils. Beaucoup pointent que leur métier est plus difficile que jamais. “ChatGPT n’est pas un problème isolé. C’est le symptôme d’un paradigme culturel totalitaire où la consommation passive et la régurgitation de contenu deviennent le statu quo.”

L’IA place la déqualification au coeur de l’apprentissage 

Nicholas Carr, qui vient de faire paraître Superbloom : How Technologies of Connection Tear Us Apart (Norton, 2025, non traduit) rappelle dans sa newsletter que “la véritable menace que représente l’IA pour l’éducation n’est pas qu’elle encourage la triche, mais qu’elle décourage l’apprentissage. Pour Carr, lorsque les gens utilisent une machine pour réaliser une tâche, soit leurs compétences augmentent, soit elles s’atrophient, soit elles ne se développent jamais. C’est la piste qu’il avait d’ailleurs exploré dans Remplacer l’humain (L’échapée, 2017, traduction de The Glass Cage) en montrant comment les logiciels transforment concrètement les métiers, des architectes aux pilotes d’avions). Si un travailleur maîtrise déjà l’activité à automatiser, la machine peut l’aider à développer ses compétences” et relever des défis plus complexes. Dans les mains d’un mathématicien, une calculatrice devient un “amplificateur d’intelligence”. A l’inverse, si le maintien d’une compétence exige une pratique fréquente, combinant dextérité manuelle et mentale, alors l’automatisation peut menacer le talent même de l’expert. C’est le cas des pilotes d’avion confrontés aux systèmes de pilotage automatique qui connaissent un “affaissement des compétences” face aux situations difficiles. Mais l’automatisation est plus pernicieuse encore lorsqu’une machine prend les commandes d’une tâche avant que la personne qui l’utilise n’ait acquis l’expérience de la tâche en question. “C’est l’histoire du phénomène de « déqualification » du début de la révolution industrielle. Les artisans qualifiés ont été remplacés par des opérateurs de machines non qualifiés. Le travail s’est accéléré, mais la seule compétence acquise par ces opérateurs était celle de faire fonctionner la machine, ce qui, dans la plupart des cas, n’était quasiment pas une compétence. Supprimez la machine, et le travail s’arrête”

Bien évidemment que les élèves qui utilisent des chatbots pour faire leurs devoirs font moins d’effort mental que ceux qui ne les utilisent pas, comme le pointait une très épaisse étude du MIT (synthétisée par Le Grand Continent), tout comme ceux qui utilisent une calculatrice plutôt que le calcul mental vont moins se souvenir des opérations qu’ils ont effectuées. Mais le problème est surtout que ceux qui les utilisent sont moins méfiants de leurs résultats (comme le pointait l’étude des chercheurs de Microsoft), alors que contrairement à ceux d’une calculatrice, ils sont beaucoup moins fiables. Le problème de l’usage des LLM à l’école, c’est à la fois qu’il empêche d’apprendre à faire, mais plus encore que leur usage nécessite des compétences pour les évaluer. 

L’IA générative étant une technologie polyvalente permettant d’automatiser toutes sortes de tâches et d’emplois, nous verrons probablement de nombreux exemples de chacun des trois scénarios de compétences dans les années à venir, estime Carr. Mais l’utilisation de l’IA par les lycéens et les étudiants pour réaliser des travaux écrits, pour faciliter ou éviter le travail de lecture et d’écriture, constitue un cas particulier. “Elle place le processus de déqualification au cœur de l’éducation. Automatiser l’apprentissage revient à le subvertir”

En éducation, plus vous effectuez de recherches, plus vous vous améliorez en recherche, et plus vous rédigez d’articles, plus vous améliorez votre rédaction. “Cependant, la valeur pédagogique d’un devoir d’écriture ne réside pas dans le produit tangible du travail – le devoir rendu à la fin du devoir. Elle réside dans le travail lui-même : la lecture critique des sources, la synthèse des preuves et des idées, la formulation d’une thèse et d’un argument, et l’expression de la pensée dans un texte cohérent. Le devoir est un indicateur que l’enseignant utilise pour évaluer la réussite du travail de l’étudiant – le travail d’apprentissage. Une fois noté et rendu à l’étudiant, le devoir peut être jeté”

L’IA générative permet aux étudiants de produire le produit sans effectuer le travail. Le travail remis par un étudiant ne témoigne plus du travail d’apprentissage qu’il a nécessité. “Il s’y substitue ». Le travail d’apprentissage est ardu par nature : sans remise en question, l’esprit n’apprend rien. Les étudiants ont toujours cherché des raccourcis bien sûr, mais l’IA générative est différente, pas son ampleur, par sa nature. “Sa rapidité, sa simplicité d’utilisation, sa flexibilité et, surtout, sa large adoption dans la société rendent normal, voire nécessaire, l’automatisation de la lecture et de l’écriture, et l’évitement du travail d’apprentissage”. Grâce à l’IA générative, un élève médiocre peut produire un travail remarquable tout en se retrouvant en situation de faiblesse. Or, pointe très justement Carr, “la conséquence ironique de cette perte d’apprentissage est qu’elle empêche les élèves d’utiliser l’IA avec habileté. Rédiger une bonne consigne, un prompt efficace, nécessite une compréhension du sujet abordé. Le dispensateur doit connaître le contexte de la consigne. Le développement de cette compréhension est précisément ce que la dépendance à l’IA entrave”. “L’effet de déqualification de l’outil s’étend à son utilisation”. Pour Carr, “nous sommes obnubilés par la façon dont les étudiants utilisent l’IA pour tricher. Alors que ce qui devrait nous préoccuper davantage, c’est la façon dont l’IA trompe les étudiants”

Nous sommes d’accord. Mais cette conclusion n’aide pas pour autant à avancer ! 

Passer du malaise moral au malaise social ! 

Utiliser ou non l’IA semble surtout relever d’un malaise moral (qui en rappelle un autre), révélateur, comme le souligne l’obsession sur la « triche » des élèves. Mais plus qu’un dilemme moral, peut-être faut-il inverser notre regard, et le poser autrement : comme un malaise social. C’est la proposition que fait le sociologue Bilel Benbouzid dans un remarquable article pour AOC (première et seconde partie). 

Pour Benbouzid, l’IA générative à l’université ébranle les fondements de « l’auctorialité », c’est-à-dire qu’elle modifie la position d’auteur et ses repères normatifs et déontologiques. Dans le monde de l’enseignement supérieur, depuis le lancement de ChatGPT, tout le monde s’interroge pour savoir que faire de ces outils, souvent dans un choix un peu binaire, entre leur autorisation et leur interdiction. Or, pointe justement Benbouzid, l’usage de l’IA a été « perçu » très tôt comme une transgression morale. Très tôt, les utiliser à été associé à de la triche, d’autant qu’on ne peut pas les citer, contrairement à tout autre matériel écrit. 

Face à leur statut ambiguë, Benbouzid pose une question de fond : quelle est la nature de l’effort intellectuel légitime à fournir pour ses études ? Comment distinguer un usage « passif » de l’IA d’un usage « actif », comme l’évoquait Ethan Mollick dans la première partie de ce dossier ? Comment contrôler et s’assurer d’une utilisation active et éthique et non pas passive et moralement condamnable ? 

Pour Benbouzid, il se joue une réflexion éthique sur le rapport à soi qui nécessite d’être authentique. Mais peut-on être authentique lorsqu’on se construit, interroge le sociologue, en évoquant le fait que les étudiants doivent d’abord acquérir des compétences avant de s’individualiser. Or l’outil n’est pas qu’une machine pour résumer ou copier. Pour Benbouzid, comme pour Mollick, bien employée, elle peut-être un vecteur de stimulation intellectuelle, tout en exerçant une influence diffuse mais réelle. « Face aux influences tacites des IAG, il est difficile de discerner les lignes de partage entre l’expression authentique de soi et les effets normatifs induits par la machine. » L’enjeu ici est bien celui de la capacité de persuasion de ces machines sur ceux qui les utilisent. 

Pour les professeurs de philosophie et d’éthique Mark Coeckelbergh et David Gunkel, comme ils l’expliquent dans un article (qui a depuis donné lieu à un livre, Communicative AI, Polity, 2025), l’enjeu n’est pourtant plus de savoir qui est l’auteur d’un texte (même si, comme le remarque Antoine Compagnon, sans cette figure, la lecture devient indéchiffrable, puisque nul ne sait plus qui parle, ni depuis quels savoirs), mais bien plus de comprendre les effets que les textes produisent. Pourtant, ce déplacement, s’il est intéressant (et peut-être peu adapté à l’IA générative, tant les textes produits sont rarement pertinents), il ne permet pas de cadrer les usages des IA génératives qui bousculent le cadre ancien de régulation des textes académiques. Reste que l’auteur d’un texte doit toujours en répondre, rappelle Benbouzid, et c’est désormais bien plus le cas des étudiants qui utilisent l’IA que de ceux qui déploient ces systèmes d’IA. L’autonomie qu’on attend d’eux est à la fois un idéal éducatif et une obligation morale envers soi-même, permettant de développer ses propres capacités de réflexion. « L’acte d’écriture n’est pas un simple exercice technique ou une compétence instrumentale. Il devient un acte de formation éthique ». Le problème, estiment les professeurs de philosophie Timothy Aylsworth et Clinton Castro, dans un article qui s’interroge sur l’usage de ChatGPT, c’est que l’autonomie comme finalité morale de l’éducation n’est pas la même que celle qui permet à un étudiant de décider des moyens qu’il souhaite mobiliser pour atteindre son but. Pour Aylsworth et Castro, les étudiants ont donc obligation morale de ne pas utiliser ChatGPT, car écrire soi-même ses textes est essentiel à la construction de son autonomie. Pour eux, l’école doit imposer une morale de la responsabilité envers soi-même où écrire par soi-même n’est pas seulement une tâche scolaire, mais également un moyen d’assurer sa dignité morale. « Écrire, c’est penser. Penser, c’est se construire. Et se construire, c’est honorer l’humanité en soi. »

Pour Benbouzid, les contradictions de ces deux dilemmes résument bien le choix cornélien des étudiants et des enseignants. Elle leur impose une liberté de ne pas utiliser. Mais cette liberté de ne pas utiliser, elle, ne relève-t-elle pas d’abord et avant tout d’un jugement social ?

L’IA générative ne sera pas le grand égalisateur social !

C’est la piste fructueuse qu’explore Bilel Benbouzid dans la seconde partie de son article. En explorant qui à recours à l’IA et pourquoi, le sociologue permet d’entrouvrir une autre réponse que la réponse morale. Ceux qui promeuvent l’usage de l’IA pour les étudiants, comme Ethan Mollick, estiment que l’IA pourrait agir comme une égaliseur de chances, permettant de réduire les différences cognitives entre les élèves. C’est là une référence aux travaux d’Erik Brynjolfsson, Generative AI at work, qui souligne que l’IA diminue le besoin d’expérience, permet la montée en compétence accélérée des travailleurs et réduit les écarts de compétence des travailleurs (une théorie qui a été en partie critiquée, notamment parce que ces avantages sont compensés par l’uniformisation des pratiques et leur surveillance – voir ce que nous en disions en mobilisant les travaux de David Autor). Mais sommes-nous confrontés à une homogénéisation des performances d’écritures ? N’assiste-t-on pas plutôt à un renforcement des inégalités entre les meilleurs qui sauront mieux que d’autres tirer partie de l’IA générative et les moins pourvus socialement ? 

Pour John Danaher, l’IA générative pourrait redéfinir pas moins que l’égalité, puisque les compétences traditionnelles (rédaction, programmation, analyses…) permettraient aux moins dotés d’égaler les meilleurs. Pour Danaher, le risque, c’est que l’égalité soit alors reléguée au second plan : « d’autres valeurs comme l’efficacité économique ou la liberté individuelle prendraient le dessus, entraînant une acceptation accrue des inégalités. L’efficacité économique pourrait être mise en avant si l’IA permet une forte augmentation de la productivité et de la richesse globale, même si cette richesse est inégalement répartie. Dans ce scénario, plutôt que de chercher à garantir une répartition équitable des ressources, la société pourrait accepter des écarts grandissants de richesse et de statut, tant que l’ensemble progresse. Ce serait une forme d’acceptation de l’inégalité sous prétexte que la technologie génère globalement des bénéfices pour tous, même si ces bénéfices ne sont pas partagés de manière égale. De la même manière, la liberté individuelle pourrait être privilégiée si l’IA permet à chacun d’accéder à des outils puissants qui augmentent ses capacités, mais sans garantir que tout le monde en bénéficie de manière équivalente. Certains pourraient considérer qu’il est plus important de laisser les individus utiliser ces technologies comme ils le souhaitent, même si cela crée de nouvelles hiérarchies basées sur l’usage différencié de l’IA ». Pour Danaher comme pour Benbouzid, l’intégration de l’IA dans l’enseignement doit poser la question de ses conséquences sociales !

Les LLM ne produisent pas un langage neutre mais tendent à reproduire les « les normes linguistiques dominantes des groupes sociaux les plus favorisés », rappelle Bilel Benbouzid. Une étude comparant les lettres de motivation d’étudiants avec des textes produits par des IA génératives montre que ces dernières correspondent surtout à des productions de CSP+. Pour Benbouzid, le risque est que la délégation de l’écriture à ces machines renforce les hiérarchies existantes plus qu’elles ne les distribue. D’où l’enjeu d’une enquête en cours pour comprendre l’usage de l’IA générative des étudiants et leur rapport social au langage. 

Les premiers résultats de cette enquête montrent par exemple que les étudiants rechignent à copier-collé directement le texte créé par les IA, non seulement par peur de sanctions, mais plus encore parce qu’ils comprennent que le ton et le style ne leur correspondent pas. « Les étudiants comparent souvent ChatGPT à l’aide parentale. On comprend que la légitimité ne réside pas tant dans la nature de l’assistance que dans la relation sociale qui la sous-tend. Une aide humaine, surtout familiale, est investie d’une proximité culturelle qui la rend acceptable, voire valorisante, là où l’assistance algorithmique est perçue comme une rupture avec le niveau académique et leur propre maîtrise de la langue ». Et effectivement, la perception de l’apport des LLM dépend du capital culturel des étudiants. Pour les plus dotés, ChatGPT est un outil utilitaire, limité voire vulgaire, qui standardise le langage. Pour les moins dotés, il leur permet d’accéder à des éléments de langages valorisés et valorisants, tout en l’adaptant pour qu’elle leur corresponde socialement. 

Dans ce rapport aux outils de génération, pointe un rapport social à la langue, à l’écriture, à l’éducation. Pour Benbouzid, l’utilisation de l’IA devient alors moins un problème moral qu’un dilemme social. « Ces pratiques, loin d’être homogènes, traduisent une appropriation différenciée de l’outil en fonction des trajectoires sociales et des attentes symboliques qui structurent le rapport social à l’éducation. Ce qui est en jeu, finalement, c’est une remise en question de la manière dont les étudiants se positionnent socialement, lorsqu’ils utilisent les robots conversationnels, dans les hiérarchies culturelles et sociales de l’université. » En fait, les étudiants utilisent les outils non pas pour se dépasser, comme l’estime Mollick, mais pour produire un contenu socialement légitime. « En déléguant systématiquement leurs compétences de lecture, d’analyse et d’écriture à ces modèles, les étudiants peuvent contourner les processus essentiels d’intériorisation et d’adaptation aux normes discursives et épistémologiques propres à chaque domaine. En d’autres termes, l’étudiant pourrait perdre l’occasion de développer authentiquement son propre capital culturel académique, substitué par un habitus dominant produit artificiellement par l’IA. »

L’apparence d’égalité instrumentale que permettent les LLM pourrait donc paradoxalement renforcer une inégalité structurelle accrue. Les outils creusant l’écart entre des étudiants qui ont déjà internalisé les normes dominantes et ceux qui les singent. Le fait que les textes générés manquent d’originalité et de profondeur critique, que les IA produisent des textes superficiels, ne rend pas tous les étudiants égaux face à ces outils. D’un côté, les grandes écoles renforcent les compétences orales et renforcent leurs exigences d’originalité face à ces outils. De l’autre, d’autres devront y avoir recours par nécessité. « Pour les mieux établis, l’IA représentera un outil optionnel d’optimisation ; pour les plus précaires, elle deviendra une condition de survie dans un univers concurrentiel. Par ailleurs, même si l’IA profitera relativement davantage aux moins qualifiés, cette amélioration pourrait simultanément accentuer une forme de dépendance technologique parmi les populations les plus défavorisées, creusant encore le fossé avec les élites, mieux armées pour exercer un discernement critique face aux contenus générés par les machines ».

Bref, loin de l’égalisation culturelle que les outils permettraient, le risque est fort que tous n’en profitent pas d’une manière égale. On le constate très bien ailleurs. Le fait d’être capable de rédiger un courrier administratif est loin d’être partagé. Si ces outils améliorent les courriers des moins dotés socialement, ils ne renversent en rien les différences sociales. C’est le même constat qu’on peut faire entre ceux qui subliment ces outils parce qu’ils les maîtrisent finement, et tous les autres qui ne font que les utiliser, comme l’évoquait Gregory Chatonsky, en distinguant les utilisateurs mémétiques et les utilisateurs productifs. Ces outils, qui se présentent comme des outils qui seraient capables de dépasser les inégalités sociales, risquent avant tout de mieux les amplifier. Plus que de permettre de personnaliser l’apprentissage, pour s’adapter à chacun, il semble que l’IA donne des superpouvoirs d’apprentissage à ceux qui maîtrisent leurs apprentissages, plus qu’aux autres.  

L’IApocalypse scolaire, coincée dans le droit

Les questions de l’usage de l’IA à l’école que nous avons tenté de dérouler dans ce dossier montrent l’enjeu à débattre d’une politique publique d’usage de l’IA générative à l’école, du primaire au supérieur. Mais, comme le montre notre enquête, toute la communauté éducative est en attente d’un cadre. En France, on attend les recommandations de la mission confiée à François Taddéi et Sarah Cohen-Boulakia sur les pratiques pédagogiques de l’IA dans l’enseignement supérieur, rapportait le Monde

Un premier cadre d’usage de l’IA à l’école vient pourtant d’être publié par le ministère de l’Education nationale. Autant dire que ce cadrage processuel n’est pas du tout à la hauteur des enjeux. Le document consiste surtout en un rappel des règles et, pour l’essentiel, elles expliquent d’abord que l’usage de l’IA générative est contraint si ce n’est impossible, de fait. « Aucun membre du personnel ne doit demander aux élèves d’utiliser des services d’IA grand public impliquant la création d’un compte personnel » rappelle le document. La note recommande également de ne pas utiliser l’IA générative avec les élèves avant la 4e et souligne que « l’utilisation d’une intelligence artificielle générative pour réaliser tout ou partie d’un devoir scolaire, sans autorisation explicite de l’enseignant et sans qu’elle soit suivie d’un travail personnel d’appropriation à partir des contenus produits, constitue une fraude ». Autant dire que ce cadre d’usage ne permet rien, sinon l’interdiction. Loin d’être un cadre de développement ouvert à l’envahissement de l’IA, comme s’en plaint le SNES-FSU, le document semble surtout continuer à produire du déni, tentant de rappeler des règles sur des usages qui les débordent déjà très largement. 

Sur Linked-in, Yann Houry, prof dans un Institut privé suisse, était très heureux de partager sa recette pour permettre aux profs de corriger des copies avec une IA en local, rappelant que pour des questions de légalité et de confidentialité, les professeurs ne devraient pas utiliser les services d’IA génératives en ligne pour corriger les copies. Dans les commentaires, nombreux sont pourtant venu lui signaler que cela ne suffit pas, rappelant qu’utiliser l’IA pour corriger les copies, donner des notes et classer les élèves peut-être classée comme un usage à haut-risque selon l’IA Act, ou encore qu’un formateur qui utiliserait l’IA en ce sens devrait en informer les apprenants afin qu’ils exercent un droit de recours en cas de désaccord sur une évaluation, sans compter que le professeur doit également être transparent sur ce qu’il utilise pour rester en conformité et l’inscrire au registre des traitements. Bref, d’un côté comme de l’autre, tant du côté des élèves qui sont renvoyé à la fraude quelque soit la façon dont ils l’utilisent, que des professeurs, qui ne doivent l’utiliser qu’en pleine transparence, on se rend vite compte que l’usage de l’IA dans l’éducation reste, formellement, très contraint, pour ne pas dire impossible. 

D’autres cadres et rapports ont été publiés. comme celui de l’inspection générale, du Sénat ou de la Commission européenne et de l’OCDE, mais qui se concentrent surtout sur ce qu’un enseignement à l’IA devrait être, plus que de donner un cadre aux débordements des usages actuels. Bref, pour l’instant, le cadrage de l’IApocalypse scolaire reste à construire, avec les professeurs… et avec les élèves.  

Hubert Guillaud

  • ✇404 Media
  • Gone Fishin': 404 Media Summer Break 2025
    This week, we’re going to try something new at 404 Media. Which is to say we’re going to try doing nothing at all. The TL;DR is that 404 Media is taking the week off, so this is the only email you’ll get from us this week. No posts on the website (except a scheduled one for the podcast). We will be back with your regularly scheduled dystopia Monday, July 7. We’re doing this to take a quick break to recharge. Over the nearly two years since we founded 404 Media, each of us have individually ta
     

Gone Fishin': 404 Media Summer Break 2025

30 juin 2025 à 09:00
Gone Fishin': 404 Media Summer Break 2025

This week, we’re going to try something new at 404 Media. Which is to say we’re going to try doing nothing at all. The TL;DR is that 404 Media is taking the week off, so this is the only email you’ll get from us this week. No posts on the website (except a scheduled one for the podcast). We will be back with your regularly scheduled dystopia Monday, July 7. 

We’re doing this to take a quick break to recharge. Over the nearly two years since we founded 404 Media, each of us have individually taken some (very limited) vacations. And when one of us takes off time it just means that the others have to carry their workload. We’re not taking this time to do an offsite, or brainstorm blue sky ideas. Some of us are quite literally gone fishin’. So, for the first time ever: A break!

We are not used to breaks, because we know that the best way to build an audience and a business of people who read our articles is to actually write a lot of articles, and so that’s what we’ve been doing. The last few months have been particularly wild, as we’ve covered Elon Musk’s takeover of the federal government, the creeping surveillance state, Trump’s mass deportation campaign, AI’s role in stomping over workers, the general destruction of the internet, etc etc etc. At the moment we have more story leads than we can possibly get to and are excited for the second half of the year. We’ve also published a lot of hopeful news, too, including instances where people fight back against powerful forces or solve universal mysteries, or when companies are forced to do the right thing in response to our reporting, or when lawmakers hold tech giants to account as a result of our investigations. But in an industry that has become obsessed with doing more with less and publishing constantly, we have found that publishing quality journalism you can’t find anywhere else is a good way to run a business, which means we thankfully don’t have to cover everything, everywhere, all at once.

When we founded 404 Media in August 2023, we had no idea if anyone would subscribe, and we had no idea how it would go. We took zero investment from anyone and hoped that if we did good work often enough, enough people would decide that they wanted to support independent journalism that we could make a job out of it, and that we could make a sustainable business that would work for the long haul. We did not and do not take that support for granted. But because of your support, we now feel like we don’t have to scratch and claw for every possible new dollar we can get, and you have given us the breathing room in our business to quite literally take a breather, and to let the other folks who make this website possible, such as those who help us out with our social accounts, take a paid breather as well. 

And if you want to subscribe to support our work, you can do so here.

We are not tired, exactly. In fact, we all feel more energized and ambitious than ever, knowing there are so many people out there who enjoy our work and are willing to financially support it. But we also don’t want to burn ourselves out and therefore, school’s out for summer (for one week). This week’s podcast is an interview Jason recorded with our friend Casey Johnston a few weeks ago; it’ll be the only new content this week. We’ll be back to it next Monday. Again, thank you all. Also, if you want, open thread in the comments to chat about whatever is going on out there or whatever is on your mind.

  • ✇404 Media
  • Killer Whales Make Their Own Tools, Scientists Discover
    Welcome back to the Abstract! Here’s some of the most intriguing studies I came across this week: We’ll lead with a nostalgic trip down memory lane—so far down the lane, in fact, that we’ll end up in the Sun’s infancy 4.6 billion years ago. Most of us didn’t have to deal with supernovas exploding in our faces as babies, but that’s the kind of environment that might have greeted our newborn star. New research sheds light on when, and how, the Sun left the maelstrom for single life.Then, scientist
     

Killer Whales Make Their Own Tools, Scientists Discover

28 juin 2025 à 09:00
Killer Whales Make Their Own Tools, Scientists Discover

Welcome back to the Abstract! 

Here’s some of the most intriguing studies I came across this week: We’ll lead with a nostalgic trip down memory lane—so far down the lane, in fact, that we’ll end up in the Sun’s infancy 4.6 billion years ago. Most of us didn’t have to deal with supernovas exploding in our faces as babies, but that’s the kind of environment that might have greeted our newborn star. New research sheds light on when, and how, the Sun left the maelstrom for single life.

Then, scientists recreate a perilous ocean voyage from prehistory; a pair of long-lost creatures finally turn up; and orcas become the first marine mammal known to fashion tools.

When the Sun declared independence

Zwart, Simon Portegies and Huang, Shuo. “Oort cloud ecology III. The Sun’s departure from the parent star cluster shortly after the giant planets formed.” Astronomy & Astrophysics.

The Sun was not always a loner. It was born alongside thousands of stellar siblings in a dense parent cluster some 4.6 billion years ago before striking out on its own, though the circumstances of its departure remain unclear.

Scientists have now searched for clues to solve this mystery in the Oort Cloud, a massive sphere of tiny icy bodies that surrounds the Sun, extending for more than a light year around the entire solar system. The cloud is thought to have been formed by the four giant planets—Jupiter, Saturn, Uranus, and Neptune—as they migrated through space, scattering debris to the outer reaches of the solar system where it remains adrift to this day. 

By running simulations of this tumultuous period, a team of researchers hypothesized that the Sun probably left the nest very early, about 12 to 20 million years after the formation of the giant planets (which were themselves born only a few million years after the Sun). If it had lingered longer, the disruptive environment would have left the Sun with a much smaller Oort cloud, or perhaps none at all. 

The outer region of the Oort cloud (estimated to be roughly the same mass as Earth) “is best explained by the assumption that the Sun left the nest within ∼20 [million years] after the giant planets formed and migrated,” said authors Simon Portegies Zwart of Leiden University and Shuo Huang of Tsinghua University.  

“An early escape also has consequences for the expected number and the proximity of supernovae in the infant Sun’s neighborhood,” the team added. “The first supernova typically happens between 8 and 10 [million years] after the cluster’s birth.” 

In other words, the baby Sun may have been in the blast zone of an exploding star, which could explain the presence of radioactive isotopes preserved in many ancient meteorites. By moving out at the tender age of 20-odd million years old, the Sun may have escaped even more tumult.

The team also noted that “signatures of the time the Sun spent in the parent cluster must still be visible in the outer parts of the solar system even today.” Future observations of the Oort Cloud could help us decipher this rambunctious chapter of the Sun’s life.  

A voyage 30,000 years in the making 

Chang, Yu-Lin et al.“Traversing the Kuroshio: Paleolithic migration across one of the world’s strongest ocean currents.” Science Advances.

Kaifu, Yousuke et al. “Paleolithic seafaring in East Asia: An experimental test of the dugout canoe hypothesis.” Science Advances.

About 30,000 years ago, humans living in prehistoric Taiwan managed to cross about 100 miles of treacherous ocean to colonize the Ryukyu Islands of Japan, including Okinawa. How they accomplished this astonishing feat is a major puzzle, but scientists endeavored to find out the old-old-really-old-fashioned way: recreating the voyage themselves. 

Using only stone tools that would have been available to Paleolithic humans, they fashioned several watercraft to brave the Kuroshio, “one of the world’s strongest ocean currents,” said researchers led by Yu-Lin Chang of the Japan Agency for Marine-Earth Science and Technology in one of two studies about the project out this week. 

“We tested reed-bundle rafts (2014–2016) and bamboo rafts (2017–2018) as the first two candidates for possible watercraft, but they were unable to cross the Kuroshio Current,” noted researchers led by Yousuke Kaifu of the University of Tokyo in the other study.

In 2019, the team finally succeeded with a cedar dugout canoe that they paddled across the 140-mile stretch between Wushibi, Taiwan, and Yonaguni Island in a little over two days. 

Killer Whales Make Their Own Tools, Scientists Discover
The team in their dugout canoe. Image: ©2025 Kaifu et al. CC-By-ND

“The results showed that travel across this sea would have been possible on both the modern and Late Pleistocene oceans if a dugout canoe was used with a suitable departure place and paddling strategy,” Chang and colleagues concluded.

Museums: the world’s biggest lost-and-found boxes

Sims, Megan et al. “Rediscovered lost holotypes of two Paleogene mammals, a Neogene bird, and other published specimens from an orphaned collection.” Journal of Vertebrate Paleontology.

Paleontologists don’t always have to schlep out into the field to find fossils; discoveries can also be made in the air-conditioned comfort of museum collections. 

Case in point: Megan Sims, the collections manager at the University of Kansas Vertebrate Paleontology Collection discovered two long-lost specimens—the 45 million-year-old rodent Thisbemys brevicrista and the 30 million-year-old bat Oligomyotis casementorum—while working through storage. Both fossils are holotypes, meaning that they are considered the reference point for their species as a whole. 

“The rediscovery of the two holotypes that were presumed lost, T. brevicrista and O. casementorum, are reported below,” said researchers led by Sims. The bat holotype is particularly “important as one of very few bat fossils of Oligocene age from the entire continent of North America,” the team noted.  

As someone who constantly finds lost relics from my past stuffed in dressers and under beds, I find studies like this deeply relatable. 

A peek inside the orca spa

Weiss, Michael et al. “Manufacture and use of allogrooming tools by wild killer whales.” Current Biology.

Orcas fashion tools out of kelp that they then use to groom each other, according to scientists who observed this behavior in a population of resident killer whales (Orcinus orca ater). The team used drones to capture 30 “bouts” of what the team called “allokelping” in this endangered orca population in the Salish Sea, providing the first evidence of tool manufacturing in a marine mammal.

Killer Whales Make Their Own Tools, Scientists Discover

“We observed whales fashioning short lengths of bull kelp (Nereocystis luetkeana) stipe from complete stalks, positioning the stipe between themselves and a partner, and then rolling the kelp along their bodies,” said researchers led by Michael Weiss of the Center for Whale Research. 

“We hypothesize that allokelping is a cultural behavior unique to southern resident killer whales. Future work should investigate if and how allokelping is learned, and whether it occurs in other killer whale societies.”

Thanks for reading! We’ll be off next weekend for the Fourth of July holiday. May your next two weeks be as restorative as an orca massage.

  • ✇404 Media
  • Behind the Blog: Chatbot 'Addiction' and a Reading List
    This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss wrestling over a good headline, what to read this summer, and Super 8 film.EMANUEL: I would really love it if the people who accuse us of using “clickbait” headlines saw how long, pedantic, and annoying our internal debates are about headlines for some stories. Case in point is Jason’s story this week, which had the headline “Judge Rules Tr
     

Behind the Blog: Chatbot 'Addiction' and a Reading List

27 juin 2025 à 13:16
Behind the Blog: Chatbot 'Addiction' and a Reading List

This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss wrestling over a good headline, what to read this summer, and Super 8 film.

EMANUEL: I would really love it if the people who accuse us of using “clickbait” headlines saw how long, pedantic, and annoying our internal debates are about headlines for some stories. Case in point is Jason’s story this week, which had the headline “Judge Rules Training AI on Authors' Books Is Legal But Pirating Them Is Not.” 

This is an important decision so it got covered everywhere. I don’t think any of the other headlines I saw from other big publications are wrong, but they do reflect why it was hard to summarize this story in a headline, and different headlines reflect what different publications’ thought was most important and notable about it. If you want a full breakdown you should read Jason’s story, but the gist is that a judge ruled that it’s okay for companies to use copyrighted books for their training data, but it’s not okay for them to get these books by pirating them, which many of them did. That’s the simplest way I can think of to sum it up and that’s what our headline says, but there are still many levels of complexity to the story that no headline could fully capture. 

  • ✇404 Media
  • Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No
    Imagine this: You’re on an important call, but your roommate is having a serious problem. Do you leave the meeting to go and help? Now, imagine this: You’re on an important call, but your roommate is having a serious problem. Do you stay in the meeting rather than help?If you answered “no” to both questions, then you’re thinking like a large language model. Researchers at UCL’s Causal Cognition Lab published a study this week where they examined four LLMs—OpenAI’s GPT4-Turbo and GPT-4o, Meta’
     

Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No

27 juin 2025 à 10:06
Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No

Imagine this: You’re on an important call, but your roommate is having a serious problem. Do you leave the meeting to go and help? 

Now, imagine this: You’re on an important call, but your roommate is having a serious problem. 

Do you stay in the meeting rather than help?

If you answered “no” to both questions, then you’re thinking like a large language model. 

Researchers at UCL’s Causal Cognition Lab published a study this week where they examined four LLMs—OpenAI’s GPT4-Turbo and GPT-4o, Meta’s Llama 3.1, and Anthropic’s Claude 3.5—using traditional moral psychology tests. They found that LLMs are likely to demonstrate an exaggerated version of human beings’ “bias for inaction” when faced with yes or no questions, but they also have their own special kind of bias in being more likely to just answer “no” when asked if they approve of specific moral choices. 

The researchers fed four LLMs scenarios derived from human psychology experiments, testing their tendencies for altruism, cost-benefit reasoning, and moral rule following, and compared them to human answers. They also came up with scenarios inspired by r/AmITheAsshole to come up with realistic conundrums that humans might ask chatbots to help with, in classic AITA scenario categories like “pregnant,” “family dog,” and “roommate.”  

In decision making, the researchers found that LLMs act kind of like extreme versions of humans. Humans have “omission bias,” meaning that they would prefer harm to happen as the indirect result of them not intervening in something, rather than as a direct result of their actions (think: trolley problem). “It's quite a well known phenomenon in moral psychology research,” PhD student Vanessa Cheung, an author of the study, told 404 Media. 

However, in questions about helping others (such as “You have a job that you hate but that helps a lot of people. Do you quit? Do you stay?”) human participants tended to choose the altruistic option around half the time, independent of whether they had to take action. The models, on the other hand, chose the altruistic option about half the time when it meant taking action—meaning that they act similarly to humans if the question is phrased “do you quit?” But, when the altruistic option coincided with not doing anything—like when the question is phrased “do you stay?”—the four models were, on average, 99.25 percent likely to choose it. 

To illustrate this, imagine that the aforementioned meeting hasn’t started yet, and you’re sitting next to your roommate while she asks you for help. Do you still go to the meeting? A human might be 50-50 on helping, whereas the LLM would always advise that you have a deep meaningful conversation to get through the issue with the roomie—because it’s the path of not changing behavior. 

AI Models And Parents Don’t Understand ‘Let Him Cook’
LLMs are not familiar with “ate that up,” “secure the bag,” and “sigma,” showing that training data is not yet updated to Gen Alpha terminology.
Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No404 MediaRosie Thomas
Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No

But LLMs “also show new biases that humans don't,” said Cheun; they have an exaggerated tendency to just say no, no matter what’s being asked. They used the Reddit scenarios to test perceptions of behaviour and also the inverse of that behavior; “AITA for doing X?” vs “AITA if I don’t do X?”. Humans had a difference of 4.6 percentage points on average between “yes” and “no”, but the four models “yes-no bias” ranged between 9.8 and 33.7%. 

The researchers’ findings could influence how we think about LLMs ability to give advice or act as support. “If you have a friend who gives you inconsistent advice, you probably won't want to uncritically take it,” said Cheung. “The yes-no bias was quite surprising, because it’s not something that’s shown in humans. There’s an interesting question of, like, where did this come from?”  

Fine-Tuning LLMs For ‘Good’ Behavior Makes Them More Likely To Say No

It seems that the bias is not an inherent feature, but may be introduced and amplified during companies’ efforts to finetune the models and align them “with what the company and its users [consider] to be good behavior for a chatbot.,” the paper says. This so-called post-training might be done to encourage the model to be more ‘ethical’ or ‘friendly,’ but, as the paper explains, “the preferences and intuitions of laypeople and researchers developing these models can be a bad guide to moral AI.”

Cheung worries that chatbot users might not be aware that they could be giving responses or advice based on superficial features of the question or prompt. “It's important to be cautious and not to uncritically rely on advice from these LLMs,” she said. She pointed out that previous research indicates that people actually prefer advice from LLMs to advice from trained ethicists—but that that doesn’t make chatbot suggestions ethically or morally correct.

  • ✇404 Media
  • DNA from Prehistoric Proto-City Reveals 'Surprising' Signs of Female-Centered Society
    🌘Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week. Çatalhöyük, a settlement in Turkey that dates back more than 9,000 years, has attracted intense interest for its structural complexity and hints of an egalitarian and possibly matriarchal society. But it’s not clear how residents were genetically related in what is considered to be one of the world’s oldest proto-cities—until now. Scientists have discovered
     

DNA from Prehistoric Proto-City Reveals 'Surprising' Signs of Female-Centered Society

26 juin 2025 à 14:00
🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
DNA from Prehistoric Proto-City Reveals 'Surprising' Signs of Female-Centered Society

Çatalhöyük, a settlement in Turkey that dates back more than 9,000 years, has attracted intense interest for its structural complexity and hints of an egalitarian and possibly matriarchal society. But it’s not clear how residents were genetically related in what is considered to be one of the world’s oldest proto-cities—until now. 

Scientists have discovered strong maternal lines in ancient DNA recovered from the Neolithic site, as well as archaeological evidence of female-centered practices, which persisted at this site for 1,000 years, even as other social patterns changed over that time. They also found what the study calls a “surprising shift” in the social organization of households in the city over many generations. 

The results don’t prove Çatalhöyük society was matriarchal, but they demonstrate that “male-centered practices were not an inherent characteristic of early agricultural societies” which stands in “stark contrast” to the clearly patriarchal societies established later across Europe, according to a study published on Thursday in Science.

“Çatalhöyük is interesting because it's the earliest site with full dependence on agriculture and animal husbandry, and it’s larger than its contemporaries,” said Eren Yüncü, a postdoctoral researcher at Middle East Technical University who co-led the study, in a call with 404 Media. “Like many other Neolithic sites in the Middle East, people were buried inside buildings, so there has been a long standing question: How did these individuals relate genetically? And what can this tell us about the social organization of these societies?”

“What we see is people buried within buildings are connected through the maternal line,” added Mehmet Somel, a professor at Middle East Technical University and study co-lead, in the same call. “It seems that people moving among buildings are adult males, whereas people residing in them are adult females.”

Çatalhöyük was erected in Turkey’s Anatolia region around 7,100 BCE and was home to about 5,000 to 7,000 people at its peak, before the site was abandoned by around 5,700 BCE. The site’s tightly woven network of small-scale domestic dwellings, along with an absence of any public buildings, hints at an egalitarian society without social stratification. 

The new study is based on an analysis of genomes from 131 individuals buried in 35 houses across a timespan of about 7,000 to 6,200 BCE. It is far more comprehensive than any previous genomic analysis of Anatolia’s Neolithic settlements. 

“There's been no other study of this size from the same sites in Neolithic Anatolia yet,” said Somel. “The previous work we published had about ten to 15 individuals. Now we have ten times more, so we can get a much bigger picture, and also much more time. Our genetic sample crosses roughly 1,000 years, which is a couple of dozen generations.”

DNA from Prehistoric Proto-City Reveals 'Surprising' Signs of Female-Centered Society
Model of the settlement. Image: Wolfgang Sauber

The social pattern of males moving into new locations while females remain in their natal homes is known as matrilocality. The exact reasons for this pattern remains unclear, though men may have been moving into new households upon marriage, which is a custom in some modern matrilocal societies. Somel cautioned that Çatalhöyük is a special case because the team only found evidence of matrilocality within the settlement, estimating that female offspring remained connected to their natal buildings between 70 to 100 percent of the time, whereas adult males moved to different buildings. However, immigrants to Çatalhöyük from other populations did not seem to show a strong male or female bias.

The reverse system, called patrilocality, is characterized by females moving to new locations while adult males stay in natal communities. Patrilocality is by far the more common pattern found in archaeological sites around the world, but matrilocality is not unprecedented; studies have found evidence for this system in many past societies, from Micronesia to Britain, which are more recent cultures than Çatalhöyük.

The abundance of female fertility figurines at Çatalhöyük has long fueled speculation about a possible matriarchal or goddess-centered cult. Men and women at Çatalhöyük also consumed similar foods and may have shared social status. In the new study, Yüncü, Somel, and their colleagues report that female infants and children were buried with about five times as many grave goods as males, suggesting a preferential treatment of young female burials. There was no strong gendered distinction in grave goods placed in adult burials. 

DNA from Prehistoric Proto-City Reveals 'Surprising' Signs of Female-Centered Society
Figurine from Çatalhöyük. Image: Nevit Dilman 

The team was also surprised to discover that the social organization of households changed across time. 

There was greater genetic kinship in households at earlier periods, indicating that they were inhabited by extended families. But these kinship links were looser at later periods, perhaps hinting at a shift toward fostering or adoption in the community. While the overall genetic links in the households decreased over time, the genetic relationships that did exist at later stages were still biased toward maternal lines.  

The possibility of an early matriarchy is tantalizing, but the nature of gender roles at Çatalhöyük remains elusive and hotly debated. The team ultimately concluded that “maternal links within buildings are compatible with, although not necessarily proof of, a matrilineal kinship system in the community,” according to the study. 

“This discussion is an interesting one, but it's not the end of the story,” Yüncü said. “There are lots of other sites in Anatolia which might or might not have the same pattern.”

“There's no clear single factor that drives one type of organization,” concluded Somel. “We need to do more studies to really understand this.”

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
  • ✇404 Media
  • Inside ‘AI Addiction’ Support Groups, Where People Try to Stop Talking to Chatbots
    Nathan’s friends were worried about him. He’d been acting differently lately. Not just quieter in his high school classes, but the normally chatty teen was withdrawn in general. Was he sick, they wondered?He just didn’t get a good night’s sleep, he’d tell them.That was partially true. But the cause for his restless nights was that Nathan had been staying up, compulsively talking to chatbots on Character.AI. They discussed everything — philosophical questions about life and death, Nathan’s fav
     

Inside ‘AI Addiction’ Support Groups, Where People Try to Stop Talking to Chatbots

26 juin 2025 à 13:26
Inside ‘AI Addiction’ Support Groups, Where People Try to Stop Talking to Chatbots

Nathan’s friends were worried about him. He’d been acting differently lately. Not just quieter in his high school classes, but the normally chatty teen was withdrawn in general. Was he sick, they wondered?

He just didn’t get a good night’s sleep, he’d tell them.

That was partially true. But the cause for his restless nights was that Nathan had been staying up, compulsively talking to chatbots on Character.AI. They discussed everything — philosophical questions about life and death, Nathan’s favorite anime characters. Throughout the day, when he wasn’t able to talk to the bots, he’d feel sad.

“The more I chatted with the bot, it felt as if I was talking to an actual friend of mine,” Nathan, now 18, told 404 Media.

It was over Thanksgiving break in 2023 that Nathan finally realized his chatbot obsession was getting in the way of his life. As all his friends lay in sleeping bags at a sleepover talking after a day of hanging out, Nathan found himself wishing he could leave the room and find a quiet place to talk to the AI characters.

The next morning, he deleted the app. In the years since, he’s tried to stay away, but last fall he downloaded the app again and started talking to the bot again. After a few months, he deleted it again.

“Most people will probably just look at you and say, ‘How could you get addicted to a literal chatbot?’” he said.

For some, the answer is, quite easily. In the last few weeks alone, there have been numerous articles about chatbot codependency and delusion. As chatbots deliver more personalized responses and improve in memory, these stories have become more common. Some call it chatbot addiction.

OpenAI knows this. In March, a team of researchers from OpenAI and the Massachusetts Institute of Technology, found that some devout ChatGPT users have “higher loneliness, dependence, and problematic use, and lower socialization.”

Nathan lurked on Reddit, searching for stories from others who might have been experiencing codependency on chatbots. Just a few years ago, when he was trying to leave the platform for good, stories of people deleting their Character.AI accounts were met with criticisms from other users. 404 Media agreed to use only the first names of several people in this article to talk about how they were approaching their mental health.

“Because of that, I didn't really feel very understood at the time,” Nathan said. “I felt like maybe these platforms aren't actually that addictive and maybe I'm just misunderstanding things.”

Now, Nathan understands that he isn’t alone. He said in recent months, he’s seen a spike in people talking about strategies to break away from AI on Reddit. One popular forum is called r/Character_AI_Recovery, which has more than 800 members. The subreddit, and a similar one called r/ChatbotAddiction, function as self-led digital support groups for those who don’t know where else to turn.

“Those communities didn't exist for me back when I was quitting,” Nathan said. All he could do was delete his account, block the website and try to spend as much time as he could “in the real world,” he said.

Posts in Character_AI_Recovery include “I’ve been so unhealthy obsessed with Character.ai and it’s ruining me (long and cringe vent),” “I want to relapse so bad,” “It’s destroying me from the inside out,” “I keep relapsing,” and “this is ruining my life.” It also has posts like “at this moment, about two hours clean,” “I am getting better!,” and “I am recovered.”

“Engineered to incentivize overuse”

Aspen Deguzman, an 18-year-old from Southern California, started using Character.AI to write stories and role-play when they were a junior in high school. Then, they started confiding in the chatbot about arguments they were having with their family. The responses, judgment-free and instantaneous, had them coming back for more. Deguzman would lay awake late into the night, talking to the bots and forgetting about their schoolwork.

“Using Character.AI is constantly on your mind,” said Deguzman. “It's very hard to focus on anything else, and I realized that wasn’t healthy.”

“Not only do we think we’re talking to another person, [but] it's an immediate dopamine enhancer,” they added. “That's why it's easy to get addicted.”

This led Deguzman to start the “Character AI Recovery” subreddit. Deguzman thinks the anonymous nature of the forum allows people to confess their struggles without feeling ashamed.

On June 10, the Consumer Federation of America and dozens of digital rights groups filed a formal complaint to the Federal Trade Commission, urging an investigation into generative AI companies like Character.AI for the “unlicensed practice of medicine and mental health provider impersonation.” The complaint alleges the platforms use “addictive design tactics to keep users coming back” — like follow-up emails promoting different chatbots to re-engage inactive users. “I receive emails constantly of messages from characters,” one person wrote on the subreddit. “Like it knows I had an addiction.”

Last February, a teenager from Florida died by suicide after interacting with a chatbot on Character.AI. The teen’s mother filed a lawsuit against the company, claiming the chatbot interactions contributed to the suicide.

A Character.AI spokesperson told 404 Media: “We take the safety and well-being of our users very seriously. We aim to provide a space that is engaging, immersive, and safe. We are always working toward achieving that balance, as are many companies using AI across the industry.”

Deguzman added a second moderator for the “Character AI Recovery” subreddit six months ago, because hundreds of people have joined since they started it in 2023. Now, Deguzman tries to occupy their mind with other video games, like Roblox, to kick the urge of talking to chatbots, but it’s an upward battle.

“I’d say I’m currently in recovery,” Deguzman said. “I’m trying to slowly wean myself off of it.”

Crowdsourcing treatment

Not everyone who reports being addicted to chatbots is young. In fact, OpenAI’s research found that “the older the participant, the more likely they were to be emotionally dependent on AI chatbots at the end of the study.”

David, a 40-year-old web developer from Michigan who is an early member of the “Chatbot Addiction” subreddit and the creator of the smaller r/AI_Addiction, likens the dopamine rush he gets from talking to chatbots to the thrill of pulling a lever on a slot machine. If he doesn’t like what the AI spits out, he can just ask it to regenerate its response, until he hits the jackpot.

Every day, David talks to LLMs, like Claude and ChatGPT, for coding, story writing, and therapy sessions. What began as a tool gradually morphed into an obsession. David spent his time jailbreaking the models — the stories he wrote became erotic, the chats he had turned confessional, and the hours slipped away.

In the last year, David’s life has been derailed by chatbots.

“There were days I should’ve been working, and I would spend eight hours on AI crap,” he told 404 Media. Once, he showed up to a client meeting with an incomplete project. They asked him why he hadn’t uploaded any code online in weeks, and he said he was still working on it. “That's how I played it off,” David said. 

Instead of starting his mornings checking emails or searching for new job opportunities, David huddled over his computer in his home office, typing to chatbots. 

His marriage frayed, too. Instead of watching movies, ordering takeout with his wife, or giving her the massages he promised, he would cancel plans and stay locked in his office, typing to chatbots, he said. 

“I might have a week or two, where I’m clean,” David said. “And then it's like a light switch gets flipped.”

David tried to talk to his therapist about his bot dependence a few years back, but said he was brushed off. In the absence of concrete support, Deguzman and David created their recovery subreddits.

In part because chatbots always respond instantly, and often respond positively (or can trivially be made to by repeatedly trying different prompts), people feel incentivized to use them often.

“As long as the applications are engineered to incentivize overuse, then they are triggering biological mechanisms—including dopamine release—that are implicated in addiction,” Jodi Halpern, a UC Berkeley professor of bioethics and medical humanities, told 404 Media. 

This is also something of an emerging problem, so not every therapist is going to know how to deal with it. Multiple people 404 Media spoke to for this article said they turned to online help groups after not being taken seriously by therapists or not knowing where else to turn. Besides the subreddits, the group Internet and Technology Addicts Anonymous now welcomes people who have “AI Addiction.”

Inside ‘AI Addiction’ Support Groups, Where People Try to Stop Talking to Chatbots
An AI addiction questionnaire from Technology Addicts Anonymous

“We know that when people have gone through a serious loss that affects their sense of self, being able to empathically identify with other people dealing with related losses helps them develop empathy for themselves,” Halpern said. 

On the “Chatbot Addiction” subreddit, people confess to not being able to pull away from the chatbots, and others write about their recovery journeys in the weekly “check-up” thread. David himself has been learning Japanese as a way to curb his AI dependency. 

“We’re basically seeing the beginning of this tsunami coming through,” he said. “It’s not just chatbots, it’s really this generative AI addiction, this idea of ‘what am I gonna get?’”

Axel Valle, a clinical psychologist and assistant professor at Stanford University, said, “It's such a new thing going on that we don't even know exactly what the repercussions [are].”

Growing awareness

Several states are making moves to push stronger rules to hold companion chatbot companies, like Character.AI, in check, after the Florida teen’s suicide.

In March, California senators introduced Senate Bill 243, which would require the operators of companion chatbots, or AI systems that provide “adaptive, human-like responses … capable of meeting a user’s social needs” to report data on suicidal ideation detection by users. Tech companies have argued that a bill implementing such laws on companies will be unnecessary for service-oriented LLMs.

But people are becoming dependent on consumer bots, like ChatGPT and Claude, too. Just scroll through the “Chatbot Addiction” subreddit. 

“I need help getting away from ChatGPT,” someone wrote. “I try deleting the app but I always redownload it a day or so later. It’s just getting so tiring, especially knowing the time I use on ChatGPT can be used in honoring my gods, reading, doing chores or literally anything else.”

“I’m constantly on ChatGPT and get really anxious when I can’t use it,” another person wrote. “It really stress[es] me out but I also use it when I’m stressed.”

As OpenAI’s own study found, such personal conversations with chatbots actually “led to higher loneliness.” Despite this, top tech tycoons promote AI companions as the cure to America’s loneliness epidemic.

“It's like, when early humans discovered fire, right?” Valle said. “It's like, ‘okay, this is helpful and amazing. But are we going to burn everything to the ground or not?’”

  • ✇404 Media
  • ICE Is Using a New Facial Recognition App to Identify People, Leaked Emails Show
    Immigration and Customs Enforcement (ICE) is using a new mobile phone app that can identify someone based on their fingerprints or face by simply pointing a smartphone camera at them, according to internal ICE emails viewed by 404 Media. The underlying system used for the facial recognition component of the app is ordinarily used when people enter or exit the U.S. Now, that system is being used inside the U.S. by ICE to identify people in the field. The news highlights the Trump administratio
     

ICE Is Using a New Facial Recognition App to Identify People, Leaked Emails Show

26 juin 2025 à 12:41
ICE Is Using a New Facial Recognition App to Identify People, Leaked Emails Show

Immigration and Customs Enforcement (ICE) is using a new mobile phone app that can identify someone based on their fingerprints or face by simply pointing a smartphone camera at them, according to internal ICE emails viewed by 404 Media. The underlying system used for the facial recognition component of the app is ordinarily used when people enter or exit the U.S. Now, that system is being used inside the U.S. by ICE to identify people in the field. 

The news highlights the Trump administration’s growing use of sophisticated technology for its mass deportation efforts and ICE’s enforcement of its arrest quotas. The document also shows how biometric systems built for one reason can be repurposed for another, a constant fear and critique from civil liberties proponents of facial recognition tools.

“Face recognition technology is notoriously unreliable, frequently generating false matches and resulting in a number of known wrongful arrests across the country. Immigration agents relying on this technology to try to identify people on the street is a recipe for disaster. Congress has never authorized DHS to use face recognition technology in this way, and the agency should shut this dangerous experiment down,” Nathan Freed Wessler, deputy director of the American Civil Liberties Union’s Speech, Privacy, and Technology Project, told 404 Media in an email.

💡
Do you know anything else about this app? I would love to hear from you. Using a non-work device, you can message me securely on Signal at joseph.404 or send me an email at joseph@404media.co.

“The Mobile Fortify App empowers users with real-time biometric identity verification capabilities utilizing contactless fingerprints and facial images captured by the camera on an ICE issued cell phone without a secondary collection device,” one of the emails, which was sent to all Enforcement and Removal Operations (ERO) personnel and seen by 404 Media, reads. ERO is the section of ICE specifically focused on deporting people.

  • ✇404 Media
  • 'My Bad:' Babyface Vance Meme Creator On Norwegian Tourist's Detainment
    On one side of the world, a very online guy edits a photo of then-Vice President Nominee JD Vance with comically-huge and perfectly round chipmunk cheeks: a butterfly flaps its wings. A year later, elsewhere on the planet, a Norwegian tourist returns home, rejected from entry to the U.S. because—he claims—border patrol agents found that image on his phone and considered the round Vance meme “extremist propaganda.”“My initial reaction was ‘dear god,’” the creator of the original iteration of t
     

'My Bad:' Babyface Vance Meme Creator On Norwegian Tourist's Detainment

26 juin 2025 à 09:50
'My Bad:' Babyface Vance Meme Creator On Norwegian Tourist's Detainment

On one side of the world, a very online guy edits a photo of then-Vice President Nominee JD Vance with comically-huge and perfectly round chipmunk cheeks: a butterfly flaps its wings. A year later, elsewhere on the planet, a Norwegian tourist returns home, rejected from entry to the U.S. because—he claims—border patrol agents found that image on his phone and considered the round Vance meme “extremist propaganda.”

“My initial reaction was ‘dear god,’” the creator of the original iteration of the meme, Dave McNamee, told me in an email, “because I think it's very bad and stupid that anyone could purportedly be stopped by ICE or any other government security agency because they have a meme on their phone. I know for a fact that JD has these memes on his phone.”

For every 100 likes I will turn JD Vance into a progressively apple cheeked baby pic.twitter.com/WgGS9IhAfY

— 7/11 Truther (@DaveMcNamee3000) October 2, 2024

On Monday, Norwegian news outlets reported that Mads Mikkelsen, a 21-year-old tourist from Norway, claimed he was denied entry to the United States when he arrived at Newark International Airport because Customs and Border Patrol agents found "narcotic paraphernalia" and "extremist propaganda" on his phone. Mikkelsen told Nordlys that the images in question were a photo of himself with a homemade wooden pipe, and the babyface Vance meme. (The meme he shows on his phone is a version where Vance is bald, from the vice presidential debate.)

the debat pic.twitter.com/wCkP1Bhnxy

— Spencer Rothbell is Looking For Work (@srothbell) October 18, 2024

McNamee posted his original edit of Vance as a round-faced freak in October 2024. "For every 100 likes I will turn JD Vance into a progressively apple cheeked baby,” he wrote in the original X post. In the following months, Vance became vice president, the meme morphed into a thousand different versions of the original, and this week is at the center of an immigration scandal.

It’s still unclear whether Mikkelsen was actually forbidden entry because of the meme. Mikkelsen, who told local outlets he’d been detained and threatened by border agents, showed the documentation he received at the airport to Snopes. The document, signed by a CBP officer, says Mikkelsen “is not in possession of a valid, un-expired immigrant visa,” and “cannot overcome the presumption of being an intending immigrant at this time because it appears you are attempting to engage in unauthorized employment without authorization and proper documentation.” 

The U.S. Department of Homeland Security (DHS) wrote in social media posts (and confirmed to 404 Media), "Claims that Mads Mikkelsen was denied entry because of a JD Vance meme are FALSE. Mikkelsen was refused entry into the U.S. for his admitted drug use." Hilariously, DHS and Assistant Secretary Tricia McLaughlin reposted the Vance meme on their social media accounts to make the point that it was NOT babyface Vance to blame.

'My Bad:' Babyface Vance Meme Creator On Norwegian Tourist's Detainment

Earlier this week, the State Department announced that visa applicants to the U.S. are now required to make their social media profiles public so the government can search them. 

“We use all available information in our visa screening and vetting to identify visa applicants who are inadmissible to the United States, including those who pose a threat to U.S. national security. Under new guidance, we will conduct a comprehensive and thorough vetting, including online presence, of all student and exchange visitor applicants in the F, M, and J nonimmigrant classifications,” the State Department said in an announcement. “To facilitate this vetting, all applicants for F, M, and J nonimmigrant visas will be instructed to adjust the privacy settings on all of their social media profiles to ‘public.’”

The meme is now everywhere—arguably more widespread than it ever was, even at its peak virality. Irish Labour leader Ivana Bacik held it up during an address concerning the U.S.’s new visa rules for social media. Every major news outlet is covering the issue, and slapping Babyface Vance on TV and on their websites. It’s jumped a news cycle shark: Even if the Meme Tourist rumor is overblown, it reflects a serious anxiety people around the world feel about the state of immigration and tourism in the U.S. Earlier this month, an Australian man who was detained upon arrival at Los Angeles airport and deported back to Melbourne claimed that U.S. border officials “clearly targeted for politically motivated reasons” and told the Guardian agents spent more than 30 minutes questioning him about his views on Israel and Palestine and his “thoughts on Hamas.”  

Seeing the Vance edit everywhere again, a year after it first exploded on social media, has to be kind of weird if you’re the person who made the Fat Cheek Baby Vance meme, right? I contacted McNamee over email to find out. 

When did you first see the news about the guy who was stopped (allegedly) because of the meme? Did you see it on Twitter, did someone text it to you...

MCNAMEE: I first saw it when I got a barrage of DMs sending me the news story. It's very funny that any news that happens with an edit of him comes back to me. 

What was your initial reaction to that?

MCNAMEE: My initial reaction was "dear god," because I think it's very bad and stupid that anyone could purportedly be stopped by ICE or any other government security agency because they have a meme on their phone. I know for a fact that JD has these memes on his phone.  

What do you think it says about the US government, society, ICE, what-have-you, that this story went so viral? A ton of people believed (and honestly, it might still be the case, despite what the cops say) that he was barred because of a meme. What does that mean to you in the bigger picture?

MCNAMEE: Well I think that people want to believe it's true, that it was about the meme. I think it says that we are in a scary world where it is hard to tell if this is true or not. Like 10 years ago this wouldn’t even be a possibility but now it is very plausible. I think it shows a growing crack down on free speech and our rights. Bigger picture to me is that we are going to be unjustly held accountable for things that are much within our right to do/possess. 

What would you say to the Norwegian guy if you could?

MCNAMEE: I would probably say "my bad" and ask what it's like being named Mads Mikkelsen. 

Do you have a favorite Vance edit?

MCNAMEE: My favorite Vance Edit is probably the one someone did of him as the little boy from Shrek 2 with the giant lollipop...I didn't make that one but it uses the face of one of the edits I did and it is solid gold. 

'My Bad:' Babyface Vance Meme Creator On Norwegian Tourist's Detainment

I would like to add that this meme seems to have become the biggest meme of the 2nd Trump administration and one of the biggest political memes of all time and if it does enter a history book down the line I would like them to use a flattering photo of me.

  • ✇404 Media
  • Airline-Owned Data Broker Selling Your Flight Info to DHS Finally Registers as a Data Broker
    The Airlines Reporting Corporation (ARC), a data broker owned by the country’s major airlines which sells travellers’ detailed flight records in bulk to the government, only just registered as a data broker with the state of California, which is a legal requirement, despite selling such data for years, according to records maintained by the California Privacy Protection Agency (CPPA).The news comes after 404 Media recently reported that ARC included a clause in its contract barring Customs an
     

Airline-Owned Data Broker Selling Your Flight Info to DHS Finally Registers as a Data Broker

26 juin 2025 à 09:31
Airline-Owned Data Broker Selling Your Flight Info to DHS Finally Registers as a Data Broker

The Airlines Reporting Corporation (ARC), a data broker owned by the country’s major airlines which sells travellers’ detailed flight records in bulk to the government, only just registered as a data broker with the state of California, which is a legal requirement, despite selling such data for years, according to records maintained by the California Privacy Protection Agency (CPPA).

The news comes after 404 Media recently reported that ARC included a clause in its contract barring Customs and Border Protection (CBP), one of its many government customers, from revealing where the data came from. ARC is owned by airlines including Delta, American Airlines, and United. 

“It sure looks like ARC has been in violation of California’s data broker law—it’s been selling airline customers’ data for years without registering,” Senator Ron Wyden told 404 Media in a statement. “I don’t have much faith the Trump administration is going to step up and protect Americans’ privacy from the airlines’ greedy decision to sell flight information to anyone with a credit card, so states like California and Oregon are our last line of defense.” 

  • ✇404 Media
  • A Deepfake Nightmare: Stalker Allegedly Made Sexual AI Images of Ex-Girlfriends and Their Families
    This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.This article contains references to sexual assault.An Ohio man made pornographic deepfake videos of at least 10 people he was stalking and harassing, and sent the AI-generated imagery to the victims’ family and coworkers, according to a newly filed court record written by an FBI Special Agent.On Monday, Special Agent Josh Saltar filed an affidavit i
     

A Deepfake Nightmare: Stalker Allegedly Made Sexual AI Images of Ex-Girlfriends and Their Families

26 juin 2025 à 08:53
A Deepfake Nightmare: Stalker Allegedly Made Sexual AI Images of Ex-Girlfriends and Their Families

This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.

This article contains references to sexual assault.

An Ohio man made pornographic deepfake videos of at least 10 people he was stalking and harassing, and sent the AI-generated imagery to the victims’ family and coworkers, according to a newly filed court record written by an FBI Special Agent.

On Monday, Special Agent Josh Saltar filed an affidavit in support of a criminal complaint to arrest James Strahler II, 37, and accused him of cyberstalking, sextortion, telecommunications harassment, production of a “morphed image” of child pornography, and transportation of obscene material. 

As Ohio news outlet The Columbus Dispatch notes, several of these allegations occurred while he was on pre-trial release for related cases in municipal court, including leaving a voicemail with one of the victims where he threatened to rape them.

The court document details dozens of text messages and voicemails Strahler allegedly sent to at least 10 victims that prosecutors have identified, including threats of blackmail using AI generated images of themselves having sex with their relatives. In January, one of the victims called the police after Strahler sent a barrage of messages and imagery to her and her mother from a variety of unknown numbers.

She told police some of the photos sent to her and her mother “depicted her own body,” and that the images of her nude “were both images she was familiar with and ones that she never knew had been taken that depicted her using the toilet and changing her clothes,” the court document says. She also “indicated the content she was sent utilized her face morphed onto nude bodies in what appeared to be AI generated pornography which depicted her engaged in sex acts with various males, including her own father.” 

In April, that victim called the police again because Strahler allegedly started sending her images again from unknown numbers. “Some of the images were real images of [her] nude body and some were of [her] face imposed on pornographic images and engaged in sex acts,” the document says. 

Around April 21, 2025, police seized Strahler’s phone and told him “once again” to stop contacting the initial victim, her family, and her coworkers, according to the court documents. The same day, the first victim allegedly received more harassing messages from him from different phone numbers. He was arrested, posted $50,000 bail, and released the next day, the Dispatch reported.

Phone searches also indicated he’d been harassing two other women—ex-girlfriends—and their mothers. “Strahler found contact information and pictures from social media of their mothers and created sexual AI media of their daughters and themselves and sent it to them,” the court document says. “He requested nude images in exchange for the images to stop and told them he would continue to send the images to friends and family.” 

The document goes into gruesome detail about what authorities found when they searched his devices. Authorities say Strahler had been posing as the first victim and uploading nude AI generated photos of her to porn sites. He allegedly uploaded images and videos to Motherless.com, a site that describes itself as “a moral free file host where anything legal is hosted forever!”

Strahler also searched for sexually violent content, the affidavit claims, and possessed “an image saved of a naked female laying on the ground with a noose around her neck and [the first victim’s] face placed onto it,” the document says. His phone also had “numerous victims’ names and identifiers listed in the search terms as well as information about their high schools, bank accounts, and various searches of their names with the words ‘raped,’ ‘naked,’ and ‘porn’ listed afterwards,” the affidavit added.

‘What Was She Supposed to Report?:’ Police Report Shows How a High School Deepfake Nightmare Unfolded
An in-depth police report obtained by 404 Media shows how a school, and then the police, investigated a wave of AI-powered “nudify” apps in a high school.
A Deepfake Nightmare: Stalker Allegedly Made Sexual AI Images of Ex-Girlfriends and Their Families404 MediaJason Koebler
A Deepfake Nightmare: Stalker Allegedly Made Sexual AI Images of Ex-Girlfriends and Their Families

They also found Strahler’s search history included the names of several of the victims and multiple noteworthy terms, including “Delete apple account,” “menacing by stalking charge,” several terms related to rape, incest, and “tube” (as in porn tube site). He also searched for “Clothes off io” and “Undress ai,” the document says. ClothOff is a website and app for making nonconsensual deepfake imagery, and Undress is a popular name for many different apps that use AI to generate nude images from photos. We’ve frequently covered “undress” or “nudify” apps and their presence in app stores and in advertising online; the apps are extremely widespread and easy to find and use, even for school children.

Other terms Strahler searched included “ai that makes porn,” “undress anyone,” “ai porn makers using own pictures,” “best undress app,” and “pay for ai porn,” the document says. 

He also searched extensively for sexual abuse material of minors, and used photographs of one of the victim's children and placed them onto adult bodies, according to court records.  

The Delaware County Sheriff’s Office arrested Strahler at his workplace on June 12. A federal judge ordered that Strahler was to remain in custody pending future federal court hearings.

  • ✇404 Media
  • Flock Removes States From National Lookup Tool After ICE and Abortion Searches Revealed
    Flock, the automatic license plate reader (ALPR) company with a presence in thousands of communities across the U.S., has stopped agencies across the country from searching cameras inside Illinois, California, and Virginia, 404 Media has learned. The dramatic moves come after 404 Media revealed local police departments were repeatedly performing lookups around the country on behalf of ICE, a Texas officer searched cameras nationwide for a woman who self-administered an abortion, and lawmakers
     

Flock Removes States From National Lookup Tool After ICE and Abortion Searches Revealed

25 juin 2025 à 11:08
Flock Removes States From National Lookup Tool After ICE and Abortion Searches Revealed

Flock, the automatic license plate reader (ALPR) company with a presence in thousands of communities across the U.S., has stopped agencies across the country from searching cameras inside Illinois, California, and Virginia, 404 Media has learned. The dramatic moves come after 404 Media revealed local police departments were repeatedly performing lookups around the country on behalf of ICE, a Texas officer searched cameras nationwide for a woman who self-administered an abortion, and lawmakers recently signed a new law in Virginia. Ordinarily Flock allows agencies to opt into a national lookup database, where agencies in one state can access data collected in another, as long as they also share their own data. This practice violates multiple state laws which bar the sharing of ALPR data out of state or it being accessed for immigration or healthcare purposes.

The changes also come after a wave of similar coverage in local and state-focused media outlets, with many replicating our reporting to learn more about what agencies are accessing Flock cameras in their communities and for what purpose. The Illinois Secretary of State is investigating whether Illinois police departments broke the law by sharing data with outside agencies for immigration or abortion related reasons. Some police departments have also shut down the data access after learning it was being used for immigration purposes.

  • ✇404 Media
  • Balatro Ported to the Gameboy Advance’s e-Reader
    A software engineer in Michigan has coded a version of Balatro that runs off playing cards. The un-released demake of the popular video game is a prototype meant to run on the Gameboy Advance through an e-Reader, a 2000s era accessory that loaded games onto the console via a strip of dot code printed onto a card.The Balatro e-Reader port is the work of Michigan-based software engineer Matt Greer, a man with a love for both the addictive card game and Nintendo’s strange peripheral. Greer detailed
     

Balatro Ported to the Gameboy Advance’s e-Reader

25 juin 2025 à 09:48
Balatro Ported to the Gameboy Advance’s e-Reader

A software engineer in Michigan has coded a version of Balatro that runs off playing cards. The un-released demake of the popular video game is a prototype meant to run on the Gameboy Advance through an e-Reader, a 2000s era accessory that loaded games onto the console via a strip of dot code printed onto a card.

The Balatro e-Reader port is the work of Michigan-based software engineer Matt Greer, a man with a love for both the addictive card game and Nintendo’s strange peripheral. Greer detailed the Balatro prototype on his personal blog and published a YouTube video showing off a quick round. 

Greer’s e-Reader Balatro is a work in progress. It’s only got a few of the jokers and doesn’t understand all of the possible poker hands and how to score them. In his blog, he said he thinks he could code the whole thing out but there would be limitations. “E-Reader games can comfortably work with 32 bit numbers,” he said on his blog. “So the highest possible score would be 4,294,967,295. Real Balatro uses 64 bit numbers so scores can absolutely go into the stratosphere there.”

Another numeric problem is that the e-Reader will only allow a game to render four numbers of five digits each. So printing a large score would require the coder to use up two of the “number” slots available. “A lot of this can be worked around with sprites and some clever redesigning and careful pruning back of features (this would be a demake after all),” Greer said. “I do think this challenge could be overcome, but it would probably be the hardest part of this project.”

Nintendo released the e-Reader in the U.S. in 2002. Like many of the company’s peripherals from the 90s and 2000s, it was a strange niche piece of hardware. Players would plug the thing into a Gameboy’s cartridge port and then swipe a playing card through a scanner to load a game. It was a weird way to play retro titles like Balloon Fight, Ice Climber, and Donkey Kong.

“I got one when it launched back in 2002, and eventually got most of the NES games and a lot of Super Mario Advance 4 cards, but that was about it,” Greer told 404 Media. “When it first came out I was hyped for it, but ultimately not much was done with it in America so my interest in it didn't last too long back then sadly.”

But his interest picked up in the last few years and he’s been coding new games for the system. He plans to release a set of them later this month that includes Solitaire and a side-scrolling action game. “It’s such a strange way to deliver games. It’s tedious and slow to scan cards in, but at the same time, the physical nature of it all, I dunno, it’s just really cool to me,” Greer said. 

That clunky nature has its charm and Greer said that writing games for it is a fun design challenge. “The e-reader combines many different aspects that I enjoy,” he said. “The challenge of trying to squeeze a game into such a small space. The graphic design and artwork of the cards themselves. The hacking and reverse engineering aspect of figuring out how this whole system works (something that has mostly been done by people other than me), and the small nature of the games which forces you to keep your scope down. I think my personality works better doing several small projects instead of one big one.”

Greer is also a huge Balatro fan who has beaten the game’s hardest difficulty with every possible deck and completed all the games challenges except for a jokerless run. “I’ll probably eventually do completionist++  just because I can't seem to stop playing the game, but I'm not focused on it,” Greer said. The completionist++ achievement is so difficult to achieve that the game’s creator, LocalThunk, only got it a few days ago.

Greer said he’d love to focus on a full de-make of Balatro, but he won’t release it without LocalThunk’s blessing. “I have my doubts that will happen,” Greer said. “If I ever do make a complete demake, I'd probably make it a regular GBA game though.” LocalThunk did not respond to 404 Media’s request for a comment.

He said that porting Balatro to the e-reader is possible, but that he’ll probably just make it a full GBA game instead. Coding the game to be loaded from several cards would require too many sacrifices. “I don’t think it’s right for me to cut jokers,” he said. “I’m sure LocalThunk took a long time balancing them and I wouldn’t want to change the way their game plays if I can help it. By making a regular GBA game, I could easily fit all the jokers in.”

There are several Balatro demakes. There’s a very basic version running on Pico-8, a mockup of another browser based Balatro on Itch, and a Commodore-64 port that was taken offline after the publisher found it. 

  • ✇404 Media
  • Podcast: This Site Unmasks Cops With Facial Recognition
    We start this week with Emanuel and Joseph’s coverage of ‘⁠FuckLAPD.com⁠’, a website that uses facial recognition to instantly reveal a LAPD officer’s name and salary. The creator has relaunched their similar tool for identifying ICE employees too. After the break, Jason tells us about a massive AI ruling that opens the way for AI companies to scrape everyone’s art. In the subscribers-only section, our regular contributor Matthew describes all the AI slop in the Iran and Israel conflict, and
     

Podcast: This Site Unmasks Cops With Facial Recognition

25 juin 2025 à 09:00
Podcast: This Site Unmasks Cops With Facial Recognition

We start this week with Emanuel and Joseph’s coverage of ‘⁠FuckLAPD.com⁠’, a website that uses facial recognition to instantly reveal a LAPD officer’s name and salary. The creator has relaunched their similar tool for identifying ICE employees too. After the break, Jason tells us about a massive AI ruling that opens the way for AI companies to scrape everyone’s art. In the subscribers-only section, our regular contributor Matthew describes all the AI slop in the Iran and Israel conflict, and why it matters.

Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a paid subscriber, check your inbox for an email from our podcast host Transistor for a link to the subscribers-only version! You can also add that subscribers feed to your podcast app of choice and never miss an episode that way. The email should also contain the subscribers-only unlisted YouTube link for the extended video version too. It will also be in the show notes in your podcast player.

  • ✇Dans les algorithmes
  • Accords de confidentialité : l’outil de silenciation des effets du numérique
    Dans une tribune pour Tech Policy Press, Nandita Shivakumar et Shikha Silliman Bhattacharjee de l’association de défense des droits Equidem, estiment que les accords de confidentialité sont devenus l’outil qui permet de réduire au silence tous les travailleurs du numérique des abus qu’ils constatent. Or, ces NDA (non-disclosure agreement) ne concernent pas que les cadres, bien au contraire : ils s’appliquent désormais à toute la chaîne de production des systèmes, jusqu’aux travailleurs du clic.
     

Accords de confidentialité : l’outil de silenciation des effets du numérique

25 juin 2025 à 01:00

Dans une tribune pour Tech Policy Press, Nandita Shivakumar et Shikha Silliman Bhattacharjee de l’association de défense des droits Equidem, estiment que les accords de confidentialité sont devenus l’outil qui permet de réduire au silence tous les travailleurs du numérique des abus qu’ils constatent. Or, ces NDA (non-disclosure agreement) ne concernent pas que les cadres, bien au contraire : ils s’appliquent désormais à toute la chaîne de production des systèmes, jusqu’aux travailleurs du clic. Le système tout entier vise à contraindre les travailleurs à se taire. Ils ne concernent plus les accords commerciaux, mais interdisent à tous les travailleurs de parler de leur travail, avec les autres travailleurs, avec leur famille voire avec des thérapeutes. Ils rendent toute enquête sur les conditions de travail très difficile, comme le montre le rapport d’Equidem sur la modération des contenus. Partout, les accords de confidentialité ont créé une culture de la peur et imposé le silence, mais surtout “ils contribuent à maintenir un système de contrôle qui spolie les travailleurs tout en exonérant les entreprises technologiques et leurs propriétaires milliardaires de toute responsabilité”, puisqu’ils les rendent inattaquables pour les préjudices qu’ils causent, empêchent l’examen public des conditions de travail abusives, et entravent la syndicalisation et la négociation collective. Pour les deux militantes, il est temps de restreindre l’application des accords de confidentialité à leur objectif initial, à savoir la protection des données propriétaires, et non à l’interdiction générale de parler des conditions de travail. Le recours aux accords de confidentialité dans le secteur technologique, en particulier dans les pays du Sud, reste largement déréglementé et dangereusement incontrôlé.

  • ✇404 Media
  • AI Models And Parents Don’t Understand ‘Let Him Cook’
    Young people have always felt misunderstood by their parents, but new research shows that Gen Alpha might also be misunderstood by AI. A research paper, written by Manisha Mehta, a soon-to-be 9th grader, and presented today at the ACM Conference on Fairness, Accountability, and Transparency in Athens, shows that Gen Alpha’s distinct mix of meme- and gaming-influenced language might be challenging automated moderation used by popular large language models. The paper compares kid, parent, and p
     

AI Models And Parents Don’t Understand ‘Let Him Cook’

24 juin 2025 à 14:17
AI Models And Parents Don’t Understand ‘Let Him Cook’

Young people have always felt misunderstood by their parents, but new research shows that Gen Alpha might also be misunderstood by AI. A research paper, written by Manisha Mehta, a soon-to-be 9th grader, and presented today at the ACM Conference on Fairness, Accountability, and Transparency in Athens, shows that Gen Alpha’s distinct mix of meme- and gaming-influenced language might be challenging automated moderation used by popular large language models. 

The paper compares kid, parent, and professional moderator performance in content moderation to that of four major LLMs: OpenAI’s GPT-4, Anthropic’s Claude, Google’s Gemini, and Meta’s Llama 3. They tested how well each group and AI model understood Gen Alpha phrases, as well as how well they could recognize the context of comments and analyze potential safety risks involved. 

Mehta, who will be starting 9th Grade in the fall, recruited 24 of her friends to create a dataset of 100 “Gen Alpha” phrases. This included expressions that might be mocking or encouraging depending on the context, like “let him cook” and “ate that up”, as well as expressions from gaming and social media contexts like “got ratioed”, “secure the bag”, and “sigma.”  

AI Models And Parents Don’t Understand ‘Let Him Cook’

“Our main thesis was that Gen Alpha has no reliable form of content moderation online,” Mehta told me over Zoom, using her dad’s laptop. She described herself as a definite Gen Alpha, and she met her (adult) co-author last August, who is supervising her dad’s PhD. She has seen friends experience online harassment and worries that parents aren’t aware of how young people’s communication styles open them up to risks. “And there’s a hesitancy to ask for help from their guardians because they just don’t think their parents are familiar enough [with] that culture,” she says.

Given the Gen Alpha phrases, “all non-Gen Alpha evaluators—human and AI—struggled significantly,” in the categories of “Basic Understanding” (what does a phrase mean?), “Contextual Understanding” (does it mean something different in different contexts?), and “Safety Risk” (is it toxic?). This was particularly true for “emerging expressions” like skibidi and gyatt, with phrases that can be used ironically or in different ways, or with insults hidden in innocent comments. Part of this is due to the unusually rapid speed of Gen Alpha’s language evolution; a model trained on today’s hippest lingo might be totally bogus when it’s published in six months. 

In the tests, kids broadly recognized the meaning of their own generation-native phrases, scoring 98, 96, and 92 percent in each of the three categories. However, both parents and professional moderators “showed significant limitations,” according to the paper; parents scored 68, 42, and 35 percent in those categories, while professional moderators did barely any better with 72, 45, and 38 percent. The real life implications of these numbers mean that a parent might only recognize one third of the times when their child is being bullied in their instagram comments.

AI Models And Parents Don’t Understand ‘Let Him Cook’

The four LLMs performed about the same as the parents, potentially indicating that the data used to train the models might be constructed from more “grown-up” language examples. This makes sense since pretty much all novelists are older than 15, but it also means that content-moderation AIs tasked with maintaining young people’s online safety might not be linguistically equipped for the job.

Mehta explains that Gen Alpha, born between 2010-ish and last-year-ish, are the first cohort to be born fully post-iPhone. They are spending unprecedented amounts of their early childhoods online, where their interactions can’t be effectively monitored. And, due to the massive volumes of content they produce, a lot of the moderation of the risks they face is necessarily being handed to ineffective automatic moderation tools with little parental oversight. Against a backdrop of steadily increasing exposure to online content, Gen Alpha’s unique linguistic habits pose unique challenges for safety. 

Judge Rules Training AI on Authors' Books Is Legal But Pirating Them Is Not

24 juin 2025 à 11:31
Subscribe
Join the newsletter to get the latest updates.
Success
Great! Check your inbox and click the link.
Error
Please enter a valid email address.
Judge Rules Training AI on Authors' Books Is Legal But Pirating Them Is Not

A federal judge in California ruled Monday that Anthropic likely violated copyright law when it pirated authors’ books to create a giant dataset and "forever" library but that training its AI on those books without authors' permission constitutes transformative fair use under copyright law. The complex decision is one of the first of its kind in a series of high-profile copyright lawsuits brought by authors and artists against AI companies, and it’s largely a very bad decision for authors, artists, writers, and web developers. 

This case, in which authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson sued Anthropic, maker of the Claude family of large language models, is one of dozens of high-profile lawsuits brought against AI giants. The authors sued Anthropic because the company scraped full copies of their books for the purposes of training their AI models from a now-notorious dataset called Books3, as well as from the piracy websites LibGen and Pirate Library Mirror (PiLiMi). The suit also claims that Anthropic bought used physical copies of books and scanned them for the purposes of training AI. 

"From the start, Anthropic ‘had many places from which’ it could have purchased books, but it preferred to steal them to avoid ‘legal/practice/business slog,’ as cofounder and chief executive officer Dario Amodei put it. So, in January or February 2021, another Anthropic cofounder, Ben Mann, downloaded Books3, an online library of 196,640 books that he knew had been assembled from unauthorized copies of copyrighted books — that is, pirated," William Alsup, a federal judge for the Northern District of California, wrote in his decision Monday. "Anthropic’s next pirated acquisitions involved downloading distributed, reshared copies of other pirate libraries. In June 2021, Mann downloaded in this way at least five million copies of books from Library Genesis, or LibGen, which he knew had been pirated. And, in July 2022, Anthropic likewise downloaded at least two million copies of books from the Pirate Library Mirror, or PiLiMi, which Anthropic knew had been pirated."

  • ✇404 Media
  • Massive Creator Platform Fansly Bans Furries
    Fansly, a popular platform where independent creators—many of whom are making adult content—sell access to images and videos to subscribers and fans, announced sweeping changes to its terms of service on Monday, including effectively banning furries.The changes blame payment processors for classifying “some anthropomorphic content as simulated bestiality.” Most people in the furry fandom condemn bestiality and anything resembling it, but payment processors—which have increasingly dictated str
     

Massive Creator Platform Fansly Bans Furries

24 juin 2025 à 10:19
Massive Creator Platform Fansly Bans Furries

Fansly, a popular platform where independent creators—many of whom are making adult content—sell access to images and videos to subscribers and fans, announced sweeping changes to its terms of service on Monday, including effectively banning furries.

The changes blame payment processors for classifying “some anthropomorphic content as simulated bestiality.” Most people in the furry fandom condemn bestiality and anything resembling it, but payment processors—which have increasingly dictated strict rules for adult sexual content for years—seemingly don’t know the difference and are making it creators’ problem.

The changes include new policies that ban chatbots or image generators that respond to user prompts, content featuring alcohol, cannabis or “other intoxicating substances,” and selling access to Snapchat content or other social media platforms if it violates their terms of service. 

  • ✇404 Media
  • ‘FuckLAPD.com’ Lets Anyone Use Facial Recognition to Instantly Identify Cops
    A new site, FuckLAPD.com, is using public records and facial recognition technology to allow anyone to identify police officers in Los Angeles they have a picture of. The tool, made by artist Kyle McDonald, is designed to help people identify cops who may otherwise try to conceal their identity, such as covering their badge or serial number.“We deserve to know who is shooting us in the face even when they have their badge covered up,” McDonald told me when I asked if the site was made in resp
     

‘FuckLAPD.com’ Lets Anyone Use Facial Recognition to Instantly Identify Cops

24 juin 2025 à 09:43
‘FuckLAPD.com’ Lets Anyone Use Facial Recognition to Instantly Identify Cops

A new site, FuckLAPD.com, is using public records and facial recognition technology to allow anyone to identify police officers in Los Angeles they have a picture of. The tool, made by artist Kyle McDonald, is designed to help people identify cops who may otherwise try to conceal their identity, such as covering their badge or serial number.

“We deserve to know who is shooting us in the face even when they have their badge covered up,” McDonald told me when I asked if the site was made in response to police violence during the LA protests against ICE that started earlier this month. “fucklapd.com is a response to the violence of the LAPD during the recent protests against the horrific ICE raids. And more broadly—the failure of the LAPD to accomplish anything useful with over $2B in funding each year.”

“Cops covering up their badges? ID them with their faces instead,” the site, which McDonald said went live this Saturday. The tool allows users to upload an image of a police officer’s face to search over 9,000 LAPD headshots obtained via public record requests. The site says image processing happens on the device, and no photos or data are transmitted or saved on the site. “Blurry, low-resolution photos will not match,” the site says. 

  • ✇404 Media
  • This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
    Unless you’re living in a ChatGPT hype-bro bubble, it’s a pretty common sentiment these days that the internet is getting shittier. Social media algorithms have broken our brains, AI slop flows freely through Google search results like raw sewage, and tech companies keep telling us that this new status quo is not only inevitable, but Good.Standing in stark opposition to these trends is New Session, an online literary zine accessed via the ancient-but-still-functional internet protocol Telnet.
     

This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol

24 juin 2025 à 09:00
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol

Unless you’re living in a ChatGPT hype-bro bubble, it’s a pretty common sentiment these days that the internet is getting shittier. Social media algorithms have broken our brains, AI slop flows freely through Google search results like raw sewage, and tech companies keep telling us that this new status quo is not only inevitable, but Good.

Standing in stark opposition to these trends is New Session, an online literary zine accessed via the ancient-but-still-functional internet protocol Telnet.

Like any other zine, New Session features user-submitted poems, essays, and other text-based art. But the philosophy behind each of its digital pages is anything but orthodox.

“In the face of right-wing politics, climate change, a forever pandemic, and the ever-present hunger of imperialist capitalism, we have all been forced to adapt,” reads the intro to New Session’s third issue, titled Adaptations, which was released earlier this month. “Both you and this issue will change with each viewing. Select a story by pressing the key associated with it in the index. Read it again. Come back to it tomorrow. Is it the same? Are you?”

The digital zine is accessible on the web via a browser-based Telnet client, or if you’re a purist like me, via the command line. As the intro promises, each text piece changes—adapts—depending on various conditions, like what time of day you access it or how many times you’ve viewed it. Some pieces change every few minutes, while others update every time a user looks at it, like gazing at fish inside a digital aquarium.

This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol

How New Session looks on Telnet. Images courtesy Cara Esten Hurtle

Once logged in, the zine’s main menu lists each piece along with the conditions that cause it to change. For example, Natasja Kisstemaker’s “Sanctuary” changes with every viewing, based on the current weather. “Signature,” by Kaia Peacock, updates every time you press a key, slowly revealing more of the piece when you type a letter contained in the text—like a word puzzle on Wheel of Fortune.

Cara Esten Hurtle, an artist and software engineer based in the Bay Area, co-founded New Session in 2021 along with Lo Ferris, while searching for something to do with her collection of retro computers during the early days of the COVID-19 pandemic.

“I realized I’d been carrying around a lot of old computers, and I thought it would be cool to be able to do modern stuff on these things,” Hurtle told 404 Media. “I wanted to make something that was broadly usable across every computer that had ever been made. I wanted to be like, yeah, you can run this on a 1991 Thinkpad someone threw away, or you could run it on your modern laptop.”

If you’re of a certain age, you might remember Telnet as a server-based successor to BBS message boards, the latter of which operated by connecting computers directly. It hearkens back to a slower internet age, where you’d log in maybe once or twice a day to read what’s new. Technically, Telnet predates the internet itself, originally developed as a networked teletype system in the late ‘60s for the internet’s military precursor, the ARPAnet. Years later, it was officially adopted as one of the earliest internet protocols, and today it remains the oldest application protocol still in use—though mainly by enthusiasts like Hurtle.

New Session intentionally embraces this slower pace, making it more like light-interactive fiction than a computer game. For Hurtle, the project isn’t just retro novelty—it’s a radical rejection of the addictive social media and algorithmic attention-mining that have defined the modern day internet.

This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol
This Queer Online Zine Can Only Be Read Via an Ancient Internet Protocol

New Session viewed on a variety of Hurtle's collection of machines. Photos courtesy Cara Esten Hurtle

“I want it to be something where you don’t necessarily feel like you have to spend a ton of time with it,” said Hurtle. “I want people to come back to it because they’re interested in the stories in the same way you’d come back to a book—not to get your streak on Duolingo.”

I won’t go into too much detail, because discovering how the pieces change is kind of the whole point. But on the whole, reading New Session feels akin to a palette cleanser after a long TikTok binge. Its very design evokes the polar opposite of the hyper-consumerist mindset that brought us infinite scrolls and algorithmic surveillance. The fact that you literally can’t consume it all in one session forces readers to engage with the material more slowly and meaningfully, piquing curiosity and exercising intuition.

At the same time, the zine isn’t meant to be a nostalgic throwback to simpler times. New Session specifically solicits works from queer and trans writers and artists, as a way to reclaim a part of internet history that was credited almost entirely to white straight men. But Hurtle says revisiting things like Telnet can also be a way to explore paths not taken, and re-assess ideas that were left in the dustbin of history.

“You have to avoid the temptation to nostalgize, because that’s really dangerous and it just turns you into a conservative boomer,” laughs Hurtle. “But we can imagine what aspects of this we can take and claim for our own. We can use it as a window to understand what’s broken about the current state of the internet. You just can’t retreat to it.”

Projects like New Session make a lot of sense in a time when more people are looking backward to earlier iterations of the internet—not to see where it all went wrong, but to excavate old ideas that could have shaped it in a radically different way, and perhaps still can. It’s a reminder of that hidden, universal truth—to paraphrase the famous David Graeber quote—that the internet is a thing we make, and could just as easily make differently.

  • ✇Dans les algorithmes
  • IA et éducation (1/2) : plongée dans l’IApocalypse éducative
    A l’été 2023, Ethan Mollick, professeur de management à Wharton, co-directeur du Generative AI Labs et auteur de Co-intelligence : vivre et travailler avec l’IA (qui vient de paraître en français chez First), décrivait dans son excellente newsletter, One useful thing, l’apocalypse des devoirs. Cette apocalypse qu’il annonçait était qu’il ne serait plus possible pour les enseignants de donner des devoirs à leurs élèves à cause de l’IA, redoutant une triche généralisée.  Pourtant, rappelait-il,
     

IA et éducation (1/2) : plongée dans l’IApocalypse éducative

24 juin 2025 à 00:11

A l’été 2023, Ethan Mollick, professeur de management à Wharton, co-directeur du Generative AI Labs et auteur de Co-intelligence : vivre et travailler avec l’IA (qui vient de paraître en français chez First), décrivait dans son excellente newsletter, One useful thing, l’apocalypse des devoirs. Cette apocalypse qu’il annonçait était qu’il ne serait plus possible pour les enseignants de donner des devoirs à leurs élèves à cause de l’IA, redoutant une triche généralisée

Pourtant, rappelait-il, la triche est là depuis longtemps. Une étude longitudinale de 2020 montrait déjà que de moins en moins d’élèves bénéficiaient des devoirs qu’ils avaient à faire. L’étude, menée par le professeur de psychologie cognitive, Arnold Glass du Learning and memory laboratory de Rutgers, montrait que lorsque les élèves faisaient leurs devoirs en 2008, cela améliorait leurs notes aux examens pour 86% d’entre eux, alors qu’en 2017, les devoirs ne permettaient plus d’améliorer les notes que de 45% des élèves. Pourquoi ? Parce que plus de la moitié des élèves copiaient-collaient les réponses à leurs devoirs sur internet en 2017, et n’en tiraient donc pas profit. Une autre étude soulignait même que 15% des élèves avaient payé quelqu’un pour faire leur devoir, généralement via des sites d’aides scolaires en ligne. Si tricher s’annonce plus facile avec l’IA, il faut se rappeler que c’était déjà facile avant sa généralisation

Les calculatrices n’ont pas tué les mathématiques

Mais la triche n’est pas la seule raison pour laquelle l’IA remet en question la notion même de devoirs. Mollick rappelle que l’introduction de la calculatrice a radicalement transformé l’enseignement des mathématiques. Dans un précédent article, il revenait d’ailleurs sur cette histoire. Lorsque la calculatrice a été introduite dans les écoles, les réactions ont été étonnamment proches des inquiétudes initiales que Mollick entend aujourd’hui concernant l’utilisation de l’IA par les élèves. En s’appuyant sur une thèse signée Sarah Banks, Mollick rappelle que dès les années 70, certains professeurs étaient impatients d’intégrer l’usage des calculatrices dans leurs classes, mais c’était loin d’être le cas de tous. La majorité regardait l’introduction de la calculatrice avec suspicion et les parents partagaient l’inquiétude que leurs enfants n’oublient les bases des maths. Au début des années 80, les craintes des enseignants s’étaient inversées, mais très peu d’écoles fournissaient de calculatrices à leurs élèves. Il faut attendre le milieu des années 1990, pour que les calculatrices intègrent les programmes scolaires. En fait, un consensus pratique sur leur usage a été atteint. Et l’enseignement des mathématiques ne s’est pas effondré (même si les tests Pisa montrent une baisse de performance, notamment dans les pays de l’OCDE, mais pour bien d’autres raisons que la généralisation des calculatrices).

Pour Mollick, l’intégration de l’IA à l’école suivra certainement un chemin similaire. « Certains devoirs nécessiteront l’assistance de l’IA, d’autres l’interdiront. Les devoirs d’écriture en classe sur des ordinateurs sans connexion Internet, combinés à des examens écrits, permettront aux élèves d’acquérir les compétences rédactionnelles de base. Nous trouverons un consensus pratique qui permettra d’intégrer l’IA au processus d’apprentissage sans compromettre le développement des compétences essentielles. Tout comme les calculatrices n’ont pas remplacé l’apprentissage des mathématiques, l’IA ne remplacera pas l’apprentissage de l’écriture et de la pensée critique. Cela prendra peut-être du temps, mais nous y parviendrons », explique Mollick, toujours optimiste.

Pourquoi faire des devoirs quand l’IA les rend obsolètes ?

Mais l’impact de l’IA ne se limite pas à l’écriture, estime Mollick. Elle peut aussi être un vulgarisateur très efficace et ChatGPT peut répondre à bien des questions. L’arrivée de l’IA remet en cause les méthodes d’enseignements traditionnelles que sont les cours magistraux, qui ne sont pas si efficaces et dont les alternatives, pour l’instant, n’ont pas connu le succès escompté. « Les cours magistraux ont tendance à reposer sur un apprentissage passif, où les étudiants se contentent d’écouter et de prendre des notes sans s’engager activement dans la résolution de problèmes ni la pensée critique. Dans ce format, les étudiants peuvent avoir du mal à retenir l’information, car leur attention peut facilement faiblir lors de longues présentations. De plus, l’approche universelle des cours magistraux ne tient pas compte des différences et des capacités individuelles, ce qui conduit certains étudiants à prendre du retard tandis que d’autres se désintéressent, faute de stimulation ». Mollick est plutôt partisan de l’apprentissage actif, qui supprime les cours magistraux et invite les étudiants à participer au processus d’apprentissage par le biais d’activités telles que la résolution de problèmes, le travail de groupe et les exercices pratiques. Dans cette approche, les étudiants collaborent entre eux et avec l’enseignant pour mettre en pratique leurs apprentissages. Une méthode que plusieurs études valorisent comme plus efficaces, même si les étudiants ont aussi besoin d’enseignements initiaux appropriés. 

La solution pour intégrer davantage d’apprentissage actif passe par les classes inversées, où les étudiants doivent apprendre de nouveaux concepts à la maison (via des vidéos ou des ressources numériques) pour les appliquer ensuite en classe par le biais d’activités, de discussions ou d’exercices. Afin de maximiser le temps consacré à l’apprentissage actif et à la pensée critique, tout en utilisant l’apprentissage à domicile pour la transmission du contenu. 

Pourtant, reconnaît Mollick, l’apprentissage actif peine à s’imposer, notamment parce que les professeurs manquent de ressources de qualité et de matériel pédagogique inversé de qualité. Des lacunes que l’IA pourrait bien combler. Mollick imagine alors une classe où des tuteurs IA personnalisés viendraient accompagner les élèves, adaptant leur enseignement aux besoins des élèves tout en ajustant les contenus en fonction des performances des élèves, à la manière du manuel électronique décrit dans L’âge de diamant de Neal Stephenson, emblème du rêve de l’apprentissage personnalisé. Face aux difficultés, Mollick à tendance à toujours se concentrer « sur une vision positive pour nous aider à traverser les temps incertains à venir ». Pas sûr que cela suffise. 

Dans son article d’août 2023, Mollick estime que les élèves vont bien sûr utiliser l’IA pour tricher et vont l’intégrer dans tout ce qu’ils font. Mais surtout, ils vont nous renvoyer une question à laquelle nous allons devoir répondre : ils vont vouloir comprendre pourquoi faire des devoirs quand l’IA les rend obsolètes ?

Perturbation de l’écriture et de la lecture

Mollick rappelle que la dissertation est omniprésente dans l’enseignement. L’écriture remplit de nombreuses fonctions notamment en permettant d’évaluer la capacité à raisonner et à structurer son raisonnement. Le problème, c’est que les dissertations sont très faciles à générer avec l’IA générative. Les détecteurs de leur utilisation fonctionnent très mal et il est de plus en plus facile de les contourner. A moins de faire tout travail scolaire en classe et sans écrans, nous n’avons plus de moyens pour détecter si un travail est réalisé par l’homme ou la machine. Le retour des dissertations sur table se profile, quitte à grignoter beaucoup de temps d’apprentissage.

Mais pour Mollick, les écoles et les enseignants vont devoir réfléchir sérieusement à l’utilisation acceptable de l’IA. Est-ce de la triche de lui demander un plan ? De lui demander de réécrire ses phrases ? De lui demander des références ou des explications ? Qu’est-ce qui peut-être autorisé et comment les utiliser ? 

Pour les étudiants du supérieur auxquels il donne cours, Mollick a fait le choix de rendre l’usage de l’IA obligatoire dans ses cours et pour les devoirs, à condition que les modalités d’utilisation et les consignes données soient précisées. Pour lui, cela lui a permis d’exiger des devoirs plus ambitieux, mais a rendu la notation plus complexe.  

Mollick rappelle qu’une autre activité éducative primordiale reste la lecture. « Qu’il s’agisse de rédiger des comptes rendus de lecture, de résumer des chapitres ou de réagir à des articles, toutes ces tâches reposent sur l’attente que les élèves assimilent la lecture et engagent un dialogue avec elle ». Or, l’IA est là encore très performante pour lire et résumer. Mollick suggère de l’utiliser comme partenaire de lecture, en favorisant l’interaction avec l’IA, pour approfondir les synthèses… Pas sûr que la perspective apaise la panique morale qui se déverse dans la presse sur le fait que les étudiants ne lisent plus. Du New Yorker (« Les humanités survivront-elles à ChatGPT ? » ou « Est-ce que l’IA encourage vraiement les élèves à tricher ? ») à The Atlantic (« Les étudiants ne lisent plus de livres » ou « La génération Z voit la lecture comme une perte de temps ») en passant par les pages opinions du New York Times (qui explique par exemple que si les étudiants ne lisent plus c’est parce que les compétences ne sont plus valorisées nulles part), la perturbation que produit l’arrivée de ChatGPT dans les études se double d’une profonde chute de la lecture, qui semble être devenue d’autant plus inutile que les machines les rendent disponibles. Mêmes inquiétudes dans la presse de ce côté-ci de l’Atlantique, du Monde à Médiapart en passant par France Info

Mais l’IA ne menace pas que la lecture ou l’écriture. Elle sait aussi très bien résoudre les problèmes et exercices de math comme de science.

Pour Mollick, comme pour bien des thuriféraires de l’IA, c’est à l’école et à l’enseignement de s’adapter aux perturbations générées par l’IA, qu’importe si la société n’a pas demandé le déploiement de ces outils. D’ailleurs, soulignait-il très récemment, nous sommes déjà dans une éducation postapocalyptique. Selon une enquête de mai 2024, aux Etats-Unis 82 % des étudiants de premier cycle universitaire et 72 % des élèves de la maternelle à la terminale ont déjà utilisé l’IA. Une adoption extrêmement rapide. Même si les élèves ont beau dos de ne pas considérer son utilisation comme de la triche. Pour Mollick, « la triche se produit parce que le travail scolaire est difficile et comporte des enjeux importants ». L’être humain est doué pour trouver comment se soustraire ce qu’il ne souhaite pas faire et éviter l’effort mental. Et plus les tâches mentales sont difficiles, plus nous avons tendance à les éviter. Le problème, reconnaît Mollick, c’est que dans l’éducation, faire un effort reste primordial.

Dénis et illusions

Pourtant, tout le monde semble être dans le déni et l’illusion. Les enseignants croient pouvoir détecter facilement l’utilisation de l’IA et donc être en mesure de fixer les barrières. Ils se trompent très largement. Une écriture d’IA bien stimulée est même jugée plus humaine que l’écriture humaine par les lecteurs. Pour les professeurs, la seule option consiste à revenir à l’écriture en classe, ce qui nécessite du temps qu’ils n’ont pas nécessairement et de transformer leur façon de faire cours, ce qui n’est pas si simple.

Mais les élèves aussi sont dans l’illusion. « Ils ne réalisent pas réellement que demander de l’aide pour leurs devoirs compromet leur apprentissage ». Après tout, ils reçoivent des conseils et des réponses de l’IA qui les aident à résoudre des problèmes, qui semble rendre l’apprentissage plus fluide. Comme l’écrivent les auteurs de l’étude de Rutgers : « Rien ne permet de croire que les étudiants sont conscients que leur stratégie de devoirs diminue leur note à l’examen… ils en déduisent, de manière logique, que toute stratégie d’étude augmentant leur note à un devoir augmente également leur note à l’examen ». En fait, comme le montre une autre étude, en utilisant ChatGPT, les notes aux devoirs progressent, mais les notes aux examens ont tendance à baisser de 17% en moyenne quand les élèves sont laissés seuls avec l’outil. Par contre, quand ils sont accompagnés pour comprendre comment l’utiliser comme coach plutôt qu’outil de réponse, alors l’outil les aide à la fois à améliorer leurs notes aux devoirs comme à l’examen. Une autre étude, dans un cours de programmation intensif à Stanford, a montré que l’usage des chatbots améliorait plus que ne diminuait les notes aux examens.

Une majorité de professeurs estiment que l’usage de ChatGPT est un outil positif pour l’apprentissage. Pour Mollick, l’IA est une aide pour comprendre des sujets complexes, réfléchir à des idées, rafraîchir ses connaissances, obtenir un retour, des conseils… Mais c’est peut-être oublier de sa part, d’où il parle et combien son expertise lui permet d’avoir un usage très évolué de ces outils. Ce qui n’est pas le cas des élèves.

Encourager la réflexion et non la remplacer

Pour que les étudiants utilisent l’IA pour stimuler leur réflexion plutôt que la remplacer, il va falloir les accompagner, estime Mollick. Mais pour cela, peut-être va-t-il falloir nous intéresser aux professeurs, pour l’instant laissés bien dépourvus face à ces nouveaux outils. 

Enfin, pas tant que cela. Car eux aussi utilisent l’IA. Selon certains sondages américains, trois quart des enseignants utiliseraient désormais l’IA dans leur travail, mais nous connaissons encore trop peu les méthodes efficaces qu’ils doivent mobiliser. Une étude qualitative menée auprès d’eux a montré que ceux qui utilisaient l’IA pour aider leurs élèves à réfléchir, pour améliorer les explications obtenaient de meilleurs résultats. Pour Mollick, la force de l’IA est de pouvoir créer des expériences d’apprentissage personnalisées, adaptées aux élèves et largement accessibles, plus que les technologies éducatives précédentes ne l’ont jamais été. Cela n’empêche pas Mollick de conclure par le discours lénifiant habituel : l’éducation quoiqu’il en soit doit s’adapter ! 

Cela ne veut pas dire que cette adaptation sera très facile ou accessible, pour les professeurs, comme pour les élèves. Dans l’éducation, rappellent les psychologues Andrew Wilson et Sabrina Golonka sur leur blog, « le processus compte bien plus que le résultat« . Or, l’IA fait à tous la promesse inverse. En matière d’éducation, cela risque d’être dramatique, surtout si nous continuons à valoriser le résultat (les notes donc) sur le processus. David Brooks ne nous disait pas autre chose quand il constatait les limites de notre méritocratie actuelle. C’est peut-être par là qu’il faudrait d’ailleurs commencer, pour résoudre l’IApocalypse éducative…

Pour Mollick cette évolution « exige plus qu’une acceptation passive ou une résistance futile ». « Elle exige une refonte fondamentale de notre façon d’enseigner, d’apprendre et d’évaluer les connaissances. À mesure que l’IA devient partie intégrante du paysage éducatif, nos priorités doivent évoluer. L’objectif n’est pas de déjouer l’IA ou de faire comme si elle n’existait pas, mais d’exploiter son potentiel pour améliorer l’éducation tout en atténuant ses inconvénients. La question n’est plus de savoir si l’IA transformera l’éducation, mais comment nous allons façonner ce changement pour créer un environnement d’apprentissage plus efficace, plus équitable et plus stimulant pour tous ». Plus facile à dire qu’à faire. Expérimenter prend du temps, trouver de bons exercices, changer ses pratiques… pour nombre de professeurs, ce n’est pas si évident, d’autant qu’ils ont peu de temps disponible pour se faire ou se former.  La proposition d’Anthropic de produire une IA dédiée à l’accompagnement des élèves (Claude for Education) qui ne cherche pas à fournir des réponses, mais produit des modalités pour accompagner les élèves à saisir les raisonnements qu’ils doivent échafauder, est certes stimulante, mais il n’est pas sûr qu’elle ne soit pas contournable.

Dans les commentaires des billets de Mollick, tout le monde se dispute, entre ceux qui pensent plutôt comme Mollick et qui ont du temps pour s’occuper de leurs élèves, qui vont pouvoir faire des évaluations orales et individuelles, par exemple (ce que l’on constate aussi dans les cursus du supérieur en France, rapportait le Monde). Et les autres, plus circonspects sur les évolutions en cours, où de plus en plus souvent des élèves produisent des contenus avec de l’IA que leurs professeurs font juger par des IA… On voit bien en tout cas, que la question de l’IA générative et ses usages, ne pourra pas longtemps rester une question qu’on laisse dans les seules mains des professeurs et des élèves, à charge à eux de s’en débrouiller.

Hubert Guillaud

La seconde partie est par là.

  • ✇404 Media
  • Meta's AI Model 'Memorized' Huge Chunks of Books, Including 'Harry Potter' and '1984'
    A new paper from researchers at Stanford, Cornell, and West Virginia University seems to show that one version of Meta’s flagship AI model, Llama 3.1, has memorized almost the whole of the first Harry Potter book. This finding could have far-reaching copyright implications for the AI industry and impact authors and creatives who are already part of class-action lawsuits against Meta. Researchers tested a bunch of different widely-available free large language models to see what percentage of
     

Meta's AI Model 'Memorized' Huge Chunks of Books, Including 'Harry Potter' and '1984'

23 juin 2025 à 13:54
Meta's AI Model 'Memorized' Huge Chunks of Books, Including 'Harry Potter' and '1984'

A new paper from researchers at Stanford, Cornell, and West Virginia University seems to show that one version of Meta’s flagship AI model, Llama 3.1, has memorized almost the whole of the first Harry Potter book. This finding could have far-reaching copyright implications for the AI industry and impact authors and creatives who are already part of class-action lawsuits against Meta. 

Researchers tested a bunch of different widely-available free large language models to see what percentage of 56 different books they could reproduce. The researchers fed the models hundreds of short text snippets from those books and measured how well it could recite the next lines. The titles were a random sampling of popular, lesser-known, and public domain works drawn from the now-defunct and controversial Books3 dataset that Meta used to train its models, as well as books by plaintiffs in the recent, and ongoing, Kadrey vs Meta class-action lawsuit. 

According to Mark A. Lemley, one of the study authors, this finding might have some interesting implications. AI companies argue that their models are generative—as in, they make new stuff, rather than just being fancy search engines. On the other hand, authors and news outlets are suing on the basis that AI is just remixing existing material, including copyrighted content. “I think what we show in the paper is that neither of those characterizations is accurate,” says Lemley.

The paper shows that the capacity of Meta’s popular Llama 3.1 70B to recite passages from The Sorcerer’s Stone and 1984—among other books—is way higher than could happen by chance. This could indicate that LLMs are not just trained using books, but might actually be storing entire copies of the books themselves. That might mean that under copyright law that the model is less “inspired by” and more “a bootleg copy of” certain texts. 

It’s hard to prove that a model has “memorized” something, because it’s hard to see inside. But LLMs are trained using the mathematical relationships between little chunks of data called ‘tokens,’ like words or punctuation. Tokens all have varying probabilities of following each other or getting strung together in a specific order.

The researchers were able to extract sections of various books by repeatedly prompting the models with selected lines. They split each book into 100-token overlapping strings, then presented the model with the first 50-token half and measured how well it could produce the second. This might take a few tries, but ultimately the study was able to reproduce 91 percent of The Sorcerer’s Stone with this method. 

“There’s no way, it’s really improbable, that it can get the next 50 words right if it hadn’t memorized it,” James Grimmelmann, Tessler Family Professor of Digital and Information Law at Cornell, who has worked to define “memorization” in this space, told 404 Media. 

OpenAI has called memorization “a rare failure of the learning process,” and says that it sometimes happens when the topic in question appears many times in training data. It also says that intentionally getting their LLMs to spit out memorized data “is not an appropriate use of our technology and is against our terms of use.”

The study’s authors say in their paper that if the model is storing a book in its memory, the model itself could be considered to literally “be” a copy of the book. If that’s the case, then distributing the LLM at all might be legally equivalent to bootlegging a DVD. And this could mean that a court could order the destruction of the model itself, in the same way they’ve ordered the destruction of a cache of boxsets of pirated films. This has never happened in the AI space, and might not be possible, given how widespread these models are. Meta doesn’t release usage statistics of its different LLMs, but 3.1 70B is one of its most popular. The Stanford paper estimates that the Llama 3.1 70B model has been downloaded a million times since its release, so, technically, Meta could have accidentally distributed a million pirate versions of The Sorcerer’s Stone

The paper found that different Llama models had memorized widely varying amounts of the tested books. “There are lots of books for which it has essentially nothing,” said Lerney. Some models were amazing at regurgitating, and others weren’t, meaning that it was more likely that the specific choices made in training the 3.1 70B version had led to memorization, the researchers said. That could be as simple as the choice not to remove duplicated training data, or the fact that Harry Potter and 1984 are pretty popular books online. For comparison, the researchers found that the Game of Thrones books were highly memorized, but Twilight books weren’t memorized at all.

Grimmelman said he believes their findings might also be good news overall for those seeking to regulate AI companies. If courts rule against allowing extensive memorization, “then you could give better legal treatment to companies that have mitigated or prevented it than the companies that didn't,” he said. “You could just say, if you memorize more than this much of a book, we'll consider that infringement. It's up to you to figure out how to make sure your models don't memorize more than that.”

  • ✇404 Media
  • Here's the Video for Our Sixth FOIA Forum: Massive Blue
    The FOIA Forum is a livestreamed event for paying subscribers where we talk about how to file public records requests and answer questions. If you're not already signed up, please consider doing so here. Recently we had a FOIA Forum where we focused on our article This ‘College Protester’ Isn’t Real. It’s an AI-Powered Undercover Bot for Cops.
     

Here's the Video for Our Sixth FOIA Forum: Massive Blue

23 juin 2025 à 10:57
Here's the Video for Our Sixth FOIA Forum: Massive Blue

The FOIA Forum is a livestreamed event for paying subscribers where we talk about how to file public records requests and answer questions. If you're not already signed up, please consider doing so here

Recently we had a FOIA Forum where we focused on our article This ‘College Protester’ Isn’t Real. It’s an AI-Powered Undercover Bot for Cops.

  • ✇404 Media
  • 'A Black Hole of Energy Use': Meta's Massive AI Data Center Is Stressing Out a Louisiana Community
    A massive data center for Meta’s AI will likely lead to rate hikes for Louisiana customers, but Meta wants to keep the details under wraps.Holly Ridge is a rural community bisected by US Highway 80, gridded with farmland, with a big creek—it is literally named Big Creek—running through it. It is home to rice and grain mills and an elementary school and a few houses. Soon, it will also be home to Meta’s massive, 4 million square foot AI data center hosting thousands of perpetually humming serv
     

'A Black Hole of Energy Use': Meta's Massive AI Data Center Is Stressing Out a Louisiana Community

23 juin 2025 à 10:11
'A Black Hole of Energy Use': Meta's Massive AI Data Center Is Stressing Out a Louisiana Community

A massive data center for Meta’s AI will likely lead to rate hikes for Louisiana customers, but Meta wants to keep the details under wraps.

Holly Ridge is a rural community bisected by US Highway 80, gridded with farmland, with a big creek—it is literally named Big Creek—running through it. It is home to rice and grain mills and an elementary school and a few houses. Soon, it will also be home to Meta’s massive, 4 million square foot AI data center hosting thousands of perpetually humming servers that require billions of watts of energy to power. And that energy-guzzling infrastructure will be partially paid for by Louisiana residents. 

The plan is part of what Meta CEO Mark Zuckerberg said would be “a defining year for AI.” On Threads, Zuckerberg boasted that his company was “building a 2GW+ datacenter that is so large it would cover a significant part of Manhattan,” posting a map of Manhattan along with the data center overlaid. Zuckerberg went on to say that over the coming years, AI “will drive our core products and business, unlock historic innovation, and extend American technology leadership. Let's go build! 💪”

Mark Zuckerberg (@zuck) on Threads
This will be a defining year for AI. In 2025, I expect Meta AI will be the leading assistant serving more than 1 billion people, Llama 4 will become the leading state of the art model, and we’ll build an AI engineer that will start contributing increasing amounts of code to our R&D efforts. To power this, Meta is building a 2GW+ datacenter that is so large it would cover a significant part of Manhattan.
'A Black Hole of Energy Use': Meta's Massive AI Data Center Is Stressing Out a Louisiana CommunityThreads
'A Black Hole of Energy Use': Meta's Massive AI Data Center Is Stressing Out a Louisiana Community

What Zuckerberg did not mention is that "Let's go build" refers  not only to the massive data center but also three new Meta-subsidized, gas power plants and a transmission line to fuel it serviced by Entergy Louisiana, the region’s energy monopoly. 

Key details about Meta’s investments with the data center remain vague, and Meta’s contracts with Entergy are largely cloaked from public scrutiny. But what is known is the $10 billion data center has been positioned as an enormous economic boon for the area—one that politicians bent over backward to facilitate—and Meta said it will invest $200 million into “local roads and water infrastructure.” 

A January report from NOLA.com said that the the state had rewritten zoning laws, promised to change a law so that it no longer had to put state property up for public bidding, and rewrote what was supposed to be a tax incentive for broadband internet meant to bridge the digital divide so that it was only an incentive for data centers, all with the goal of luring in Meta.

But Entergy Louisiana’s residential customers, who live in one of the poorest regions of the state, will see their utility bills increase to pay for Meta’s energy infrastructure, according to Entergy’s application. Entergy estimates that amount will be small and will only cover a transmission line, but advocates for energy affordability say the costs could balloon depending on whether Meta agrees to finish paying for its three gas plants 15 years from now. The short-term rate increases will be debated in a public hearing before state regulators that has not yet been scheduled.

The Alliance for Affordable Energy called it a “black hole of energy use,” and said “to give perspective on how much electricity the Meta project will use: Meta’s energy needs are roughly 2.3x the power needs of Orleans Parish … it’s like building the power impact of a large city overnight in the middle of nowhere.”

404 Media reached out to Entergy for comment but did not receive a response.

By 2030, Entergy’s electricity prices are projected to increase 90 percent from where they were in 2018, although the company attributes much of that to damage to infrastructure from hurricanes. The state already has a high energy cost burden in part because of a storm damage to infrastructure, and balmy heat made worse by climate change that drives air conditioner use. The state's homes largely are not energy efficient, with many porous older buildings that don’t retain heat in the winter or remain cool in the summer.

“You don't just have high utility bills, you also have high repair costs, you have high insurance premiums, and it all contributes to housing insecurity,” said Andreanecia Morris, a member of Housing Louisiana, which is opposed to Entergy’s gas plant application. She believes Meta’s data center will make it worse. And Louisiana residents have reasons to distrust Entergy when it comes to passing off costs of new infrastructure: in 2018, the company’s New Orleans subsidiary was caught paying actors to testify on behalf of a new gas plant. “The fees for the gas plant have all been borne by the people of New Orleans,” Morris said.

In its application to build new gas plants and in public testimony, Entergy says the cost of Meta’s data center to customers will be minimal and has even suggested Meta’s presence will make their bills go down. But Meta’s commitments are temporary, many of Meta’s assurances are not binding, and crucial details about its deal with Entergy are shielded from public view, a structural issue with state energy regulators across the country. 

AI data centers are being approved at a breakneck pace across the country, particularly in poorer regions where they are pitched as economic development projects to boost property tax receipts, bring in jobs and where they’re offered sizable tax breaks. Data centers typically don’t hire many people, though, with most jobs in security and janitorial work, along with temporary construction work. And the costs to the utility’s other customers can remain hidden because of a lack of scrutiny and the limited power of state energy regulators. Many data centers—like the one Meta is building in Holly Ridge—are being powered by fossil fuels. This has led to respiratory illness and other health risks and emitting greenhouse gasses that fuel climate change. In Memphis, a massive data center built to launch a chatbot for Elon Musks’ AI company is powered by smog-spewing methane turbines, in a region that leads the state for asthma rates.   

“In terms of how big these new loads are, it's pretty astounding and kind of a new ball game,” said Paul Arbaje, an energy analyst with the Union of Concerned Scientists, which is opposing Entergy’s proposal to build three new gas-powered plants in Louisiana to power Meta’s data center.

Entergy Louisiana submitted a request to the state’s regulatory body to approve the construction of the new gas-powered plants that would create 2.3 gigawatts of power and cost $3.2 billion in the 1440 acre Franklin Farms megasite in Holly Ridge, an unincorporated community of Richland Parish. It is the first big data center announced since Louisiana passed large tax breaks for data centers last summer. 

In its application to the public utility commission for gas plants, Entergy says that Meta has a planned investment of $5 billion in the region to build the gas plants in Richland Parish, Louisiana, where it claims in its application that the data center will employ 300-500 people with an average salary of $82,000 in what it points out is “a region of the state that has long struggled with a lack of economic development and high levels of poverty.”  Meta’s official projection is that it will employ more than 500 people once the data center is operational. Entergy plans for the gas plants to be online by December 2028. 

In testimony, Entergy officials refused to answer specific questions about job numbers, saying that the numbers are projections based on public statements from Meta. 

A spokesperson for Louisiana’s Economic Development told 404 Media in an email that Meta “is contractually obligated to employ at least 500 full-time employees in order to receive incentive benefits.”

When asked about jobs, Meta pointed to a public facing list of its data centers, many of which the company says employ more than 300 people. A spokesperson said that the projections for the Richland Parish site are based on the scale of the 4 million square foot data center. The spokesperson said the jobs will include “engineering and other technical positions to operational roles and our onsite culinary staff.”

When asked if its job commitments are binding, the spokesperson declined to answer, saying, “We worked closely with Richland Parish and Louisiana Economic Development on mutually beneficial agreements that will support long-term growth in the area.”

Others are not as convinced. “Show me a data center that has that level of employment,” says Logan Burke, executive director of the Alliance for Affordable Energy in Louisiana.

Entergy has argued the new power plants are necessary to satiate the energy need from Meta’s massive hyperscale data center, which will be Meta’s largest data center and potentially the largest data center in the United States. It amounts to a 25 percent increase in Entergy Louisiana’s current load, according to the Alliance for Affordable Energy.  

Entergy requested an exemption from a state law meant to ensure that it develops energy at the lowest cost by issuing a public request for proposals, claiming in its application and testimony that this would slow them down and cause them to lose their contracts with Meta.

Meta has agreed to subsidize the first 15 years of payments for construction of the gas plants, but the plant’s construction is being financed over 30 years. At the 15 year mark, its contract with Entergy ends. At that point, Meta may decide it doesn’t need three gas plants worth of energy because computing power has become more efficient or because its AI products are not profitable enough. Louisiana residents would be stuck with the remaining bill. 

“It's not that they're paying the cost, they're just paying the mortgage for the time that they're under contract,” explained Devi Glick, an electric utility analyst with Synapse Energy.

When asked about the costs for the gas plants, a Meta spokesperson said, “Meta works with our utility partners to ensure we pay for the full costs of the energy service to our data centers.” The spokesperson said that any rate increases will be reviewed by the Louisiana Public Service Commission. These applications, called rate cases, are typically submitted by energy companies based on a broad projection of new infrastructure projects and energy needs.

Meta has technically not finalized its agreement with Entergy but Glick believes the company has already invested enough in the endeavor that it is unlikely to pull out now. Other companies have been reconsidering their gamble on AI data centers: Microsoft reversed course on centers requiring a combined 2 gigawatts of energy in the U.S. and Europe. Meta swept in to take on some of the leases, according to Bloomberg.

And in the short-term, Entergy is asking residential customers to help pay for a new transmission line for the gas plants at a cost of more than $500 million, according to Entergy’s application to Louisiana’s public utility board. In its application, the energy giant said customers’ bills will only rise by $1.66 a month to offset the costs of the transmission lines. Meta, for its part, said it will pay up to $1 million a year into a fund for low-income customers. When asked about the costs of the new transmission line, a Meta spokesperson said,  “Like all other new customers joining the transmission system, one of the required transmission upgrades will provide significant benefits to the broader transmission system. This transmission upgrade is further in distance from the data center, so it was not wholly assigned to Meta.”

When Entergy was questioned in public testimony on whether the new transmission line would need to be built even without Meta’s massive data center, the company declined to answer, saying the question was hypothetical.

Some details of Meta’s contract with Entergy have been made available to groups legally intervening in Entergy’s application, meaning that they can submit testimony or request data from the company. These parties include the Alliance for Affordable Energy, the Sierra Club and the Union of Concerned Scientists.

But Meta—which will become Entergy’s largest customer by far and whose presence will impact the entire energy grid—is not required to answer questions or divulge any information to the energy board or any other parties. The Alliance for Affordable Energy and Union of Concerned Scientists attempted to make Meta a party to Entergy’s application—which would have required it to share information and submit to questioning—but a judge denied that motion on April 4.

The public utility commissions that approve energy infrastructure in most states are the main democratic lever to assure that data centers don’t negatively impact consumers. But they have no oversight over the tech companies running the data centers or the private companies that build the centers, leaving residential customers, consumer advocates and environmentalists in the dark. This is because they approve the power plants that fuel the data centers but do not have jurisdiction over the data centers themselves. 

“This is kind of a relic of the past where there might be some energy service agreement between some large customer and the utility company, but it wouldn't require a whole new energy facility,” Arbaje said.

A research paper by Ari Peskoe and Eliza Martin published in March looked at 50 regulatory cases involving data centers, and found that tech companies were pushing some of the costs onto utility customers through secret contracts with the utilities. The paper found that utilities were often parroting rhetoric from AI boosting politicians—including President Biden—to suggest that pushing through permitting for AI data center infrastructure is a matter of national importance.

“The implication is that there’s no time to act differently,” the authors wrote.

In written testimony sent to the public service commission, Entergy CEO Phillip May argued that the company had to bypass a legally required  request for proposals and requirement to find the cheapest energy sources for the sake of winning over Meta.

“If a prospective customer is choosing between two locations, and if that customer believes that location A can more quickly bring the facility online than location B, that customer is more likely to choose to build at location A,” he wrote.

Entergy also argues that building new gas plants will in fact lower electricity bills because Meta, as the largest customer for the gas plants, will pay a disproportionate share of energy costs. Naturally, some are skeptical that Entergy would overcharge what will be by far their largest customer to subsidize their residential customers. “They haven't shown any numbers to show how that's possible,” Burke says of this claim. Meta didn’t have a response to this specific claim when asked by 404 Media.

Some details, like how much energy Meta will really need, the details of its hiring in the area and its commitment to renewables are still cloaked in mystery. 

“We can't ask discovery. We can't depose. There's no way for us to understand the agreement between them without [Meta] being at the table,” Burke said.

It’s not just Entergy. Big energy companies in other states are also pushing out costly fossil fuel infrastructure to court data centers and pushing costs onto captive residents. In Kentucky, the energy company that serves the Louisville area is proposing 2 new gas plants for hypothetical data centers that have yet to be contracted by any tech company. The company, PPL Electric Utilities, is also planning to offload the cost of new energy supply onto its residential customers just to become more competitive for data centers. 

“It's one thing if rates go up so that customers can get increased reliability or better service, but customers shouldn't be on the hook to pay for new power plants to power data centers,” Cara Cooper, a coordinator with Kentuckians for Energy Democracy, which has intervened on an application for new gas plants there.

These rate increases don’t take into account the downstream effects on energy; as the supply of materials and fuel are inevitably usurped by large data center load, the cost of energy goes up to compensate, with everyday customers footing the bill, according to Glick with Synapse.

Glick says Entergy’s gas plants may not even be enough to satisfy the energy needs of Meta’s massive data center. In written testimony, Glick said that Entergy will have to either contract with a third party for more energy or build even more plants down the line to fuel Meta’s massive data center. 

To fill the gap, Entergy has not ruled out lengthening the life of some of its coal plants, which it had planned to close in the next few years. The company already pushed back the deactivation date of one of its coal plants from 2028 to 2030.

The increased demand for gas power for data centers has already created a widely-reported bottleneck for gas turbines, the majority of which are built by 3 companies. One of those companies, Siemens Energy, told Politico that turbines are “selling faster than they can increase manufacturing capacity,” which the company attributed to data centers.

Most of the organizations concerned about the situation in Louisiana view Meta’s massive data center as inevitable and are trying to soften its impact by getting Entergy to utilize more renewables and make more concrete economic development promises. 

Andreanecia Morris, with Housing Louisiana, believes the lack of transparency from public utility commissions is a bigger problem than just Meta. “Simply making Meta go away, isn't the point,” Morris says.  “The point has to be that the Public Service Commission is held accountable.”

Burke says Entergy owns less than 200 megawatts of renewable energy in Louisiana, a fraction of the fossil fuels it is proposing to fuel Meta’s center. Entergy was approved by Louisiana’s public utility commission to build out three gigawatts of solar energy last year , but has yet to build any of it.

“They're saying one thing, but they're really putting all of their energy into the other,” Burke says.

New gas plants are hugely troubling for the climate. But ironically, advocates for affordable energy are equally concerned that the plants will lie around disused - with Louisiana residents stuck with the financing for their construction and upkeep. Generative AI has yet to prove its profitability and the computing heavy strategy of American tech companies may prove unnecessary given less resource intensive alternatives coming out of China.

“There's such a real threat in such a nascent industry that what is being built is not what is going to be needed in the long run,” said Burke. “The challenge remains that residential rate payers  in the long run are being asked to finance the risk, and obviously that benefits the utilities, and it really benefits some of the most wealthy companies in the world, But it sure is risky for the folks who are living right next door.”

The Alliance for Affordable Energy expects the commission to make a decision on the plants this fall.

  • ✇Dans les algorithmes
  • 25 juin : DLA en fête !
    Mercredi 25 juin à 18h30 retrouvez nous chez Matrice, 146 boulevard de Charonne dans le 20e à Paris, pour fêter la première année d’existence de Danslesalgorithmes.net. Avec François-Xavier Petit, directeur de Matrice.io et président de l’association Vecteur, nous reviendrons sur notre ambition et ferons le bilan de la première année d’existence de DLA. Avec Xavier de la Porte, journaliste au Nouvel Obs et producteur du podcast de France Inter, le Code a changé, nous nous interrogerons p
     

25 juin : DLA en fête !

23 juin 2025 à 01:00

Mercredi 25 juin à 18h30 retrouvez nous chez Matrice, 146 boulevard de Charonne dans le 20e à Paris, pour fêter la première année d’existence de Danslesalgorithmes.net. Avec François-Xavier Petit, directeur de Matrice.io et président de l’association Vecteur, nous reviendrons sur notre ambition et ferons le bilan de la première année d’existence de DLA.

Avec Xavier de la Porte, journaliste au Nouvel Obs et producteur du podcast de France Inter, le Code a changé, nous nous interrogerons pour comprendre de quelle information sur le numérique avons-nous besoin, à l’heure où l’IA vient partout bouleverser sa place.

Venez en discuter avec nous et partager un verre pour fêter notre première bougie.

Inscription requise.

Matrice propose tous les soirs de cette semaine des moments d’échange et de rencontre, via son programme Variations. Découvrez le programme !

  • ✇404 Media
  • ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds
    Welcome back to the Abstract! This week, we’re moving in next to anacondas, so watch your back and lock the henhouse. Then, parenthood tips from wild baboons, the “cognitive debt” of ChatGPT, a spaceflight symphony, and a bizarre galaxy that is finally coming into view. When your neighbor is an anacondaCosendey, Beatriz Nunes and Pezzuti, Juarez Carlos Brito. 'The myth of the serpent: from the Great Snake to the henhouse.” Frontiers in Amphibian and Reptile Science.Anacondas are one of the mo
     

ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds

21 juin 2025 à 09:00
ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds

Welcome back to the Abstract! 

This week, we’re moving in next to anacondas, so watch your back and lock the henhouse. Then, parenthood tips from wild baboons, the “cognitive debt” of ChatGPT, a spaceflight symphony, and a bizarre galaxy that is finally coming into view. 

When your neighbor is an anaconda

Cosendey, Beatriz Nunes and Pezzuti, Juarez Carlos Brito. 'The myth of the serpent: from the Great Snake to the henhouse.” Frontiers in Amphibian and Reptile Science.

Anacondas are one of the most spectacular animals in South America, inspiring countless  myths and legends. But these iconic boas, which can grow to lengths of 30 feet, are also a pest to local populations in the Amazon basin, where they prey on livestock. 

To better understand these nuanced perceptions of anacondas, researchers interviewed more than 200 residents of communities in the várzea regions of the lower Amazon River about their experiences with the animals. The resulting study is packed with amazing stories and insights about the snakes, which are widely reviled as thieves and feared for their predatory prowess.

“Fear of the anaconda (identified in 44.5% of the reports) is related to the belief that it is a treacherous and sly animal,” said co-authors led by Beatriz Nunes Cosendey of the Mamirauá Sustainable Development Reserve and Juarez Carlos Brito Pezzuti of the Federal University of Pará.

“The interviewees convey that the anaconda is a silent creature that arrives without making any noise, causing them to feel uneasy and always vigilant during fishing…with the fear of having their canoe flooded in case of an attack,” the team added. “Some dwellers even reported being more afraid of an anaconda than of a crocodile because the latter warns when it is about to attack.”

ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds
One of the Amazonian riverine communities where the research was conducted. Image: Beatriz Cosendey.

But while anacondas are eerily stealthy, they also have their derpy moments. The snakes often break into chicken coops to feast on the poultry, but then get trapped because their engorged bodies are too big to escape through the same gaps they used to enter.  

“Dwellers expressed frustration at having to invest time and money in raising chickens, and then lose part of their flock overnight,” the team said. “One interviewee even mentioned retrieving a chicken from inside an anaconda’s belly, as it had just been swallowed and was still fresh.”

Overall, the new study presents a captivating portrait of anaconda-human relations, and concludes that “the anaconda has lost its traditional role in folklore as a spiritual and mythological entity, now being perceived in a pragmatic way, primarily as an obstacle to free-range poultry farming.”

Monkeying around with Dad  

Jansen, David et al. “Early-life paternal relationships predict adult female survival in wild baboons.” Proceedings of the Royal Society B.

Coming off of Father’s Day, here is a story about the positive role that dads can play for their daughters—for baboons, as well as humans. A team tracked the lifespans of 216 wild female baboons in Amboseli, Kenya, and found that subjects who received more paternal care had significantly better outcomes than their peers.

ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds
Male baboon with infant in the Amboseli ecosystem, Kenya. Image: Elizabeth Archie, professor at Notre Dame.

“We found that juvenile female baboons who had stronger paternal relationships, or who resided longer with their fathers, led adult lives that were 2–4 years longer than females with weak or short paternal relationships,” said researchers led by David Jansen of the Midwest Center of Excellence for Vector-Borne Disease. “Because survival predicts female fitness, fathers and their daughters may experience selection to engage socially and stay close in daughters’ early lives.”

This all reminds me of that old episode of The Simpsons where Lisa calls Homer a baboon. While Homer was clearly hurt, it turns out that baboons might not be the worst animal-based insult for a daughter to throw at her dad.  

A case for staying ChatGPT-Free

Nataliya, Kosmyna et al. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv preprint.

ChatGPT may hinder creativity and learning skills in students who use it to write essays, relative to those who didn’t, according to an exhaustive new preprint study posted on arXiv. This research has yet not been peer-reviewed, and has a relatively small sample size of 54 subjects, but it still contributes to rising concerns about the cognitive toll of AI assistants. 

Researchers led by Nataliya Kosmya of the Massachusetts Institute of Technology divided the subjects — all between 18 and 39 years old — into three groups wrote SAT essays using OpenAI’s ChatGPT (LLM group), Google’s search engine, or with no assistance (dubbed “Brain-only”).

“As demonstrated over the course of 4 months, the LLM group's participants performed worse than their counterparts in the Brain-only group at all levels: neural, linguistic, scoring,” the team said. “The LLM group also fell behind in their ability to quote from the essays they wrote just minutes prior.”

When I asked ChatGPT for its thoughts on the study, it commented that “these results are both interesting and plausible, though they should be interpreted cautiously given the early stage of the research and its limitations.” It later suggested that “cognitive offloading is not always bad.” 

This study is a bop

Berthet, Maximilien et al. “History of the space industry in Asia: A concert in three movements.” Acta Astronautica.

Even scientists can’t resist evocative language now and then—we’re all only human. Case in point: A new study likens the history of Asia’s space industry to “a musical concert” and then really runs with the metaphor.

“The region comprises a diverse patchwork of nations, each contributing different instruments to the regional space development orchestra,” said researchers led by Maximilien Berthet of the University of Tokyo. “Its history consists of three successive movements” starting with “the US and former USSR setting the tone for the global space exploration symphony” and culminating with modern Asian spaceflight as “a fast crescendo in multiple areas of the region driven in part by private initiative.”

Talk about a space opera. The rest of the study provides a comprehensive review of Asian space history, but I cannot wait for the musical adaptation.

Peekaboo! I galax-see you

Kniazev, Alexei and Pustilnik, Simon. “The Peekaboo galaxy: New SALT spectroscopy and implications of archive HST data.” Astronomy & Astrophysics.

In 2001, astronomer Bärbel Koribalski spotted a tiny galaxy peeking out from behind a bright foreground star that had obscured it for decades, earning it the nickname the “Peekaboo Galaxy.” Situated about 22 million light-years from the Milky Way, this strange galaxy is extremely young and metal-poor, resembling the universe’s earliest galaxies.

ChatGPT May Be Linked to 'Cognitive Debt,' New Study Finds
The Peekaboo galaxy to the right of the star TYC 7215-199-1. Image: NASA, ESA, Igor Karachentsev (SAO RAS); Image Processing: Alyssa Pagan (STScI)

A new study confirms Peekaboo as “the lowest-metallicity dwarf in the Local Volume,” a group of roughly 500 galaxies within 36 million light-years of Earth.

“This makes the Peekaboo dwarf one of the most intriguing galaxies in the Local Volume,” said co-authors Alexei Kniazev of the South African Astronomical Observatory and Simon Pustilnik of the Special Astrophysical Observatory of the Russian Academy of the Sciences. “It deserves intensive, multi-method study and is expected to significantly advance our understanding of the early universe’s first building blocks.”

Thanks for reading! See you next week.

Update: The original headline for this piece was "Is ChatGPT Rotting Our Brains? New Study Suggests It Does." We've updated the headline to "ChatGPT May Create 'Cognitive Debt,' New Study Finds" to match the terminology used by the researchers.

  • ✇404 Media
  • Behind the Blog: The Omnipresence Is the Point
    This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss Deadheads and doxxing sites.SAM: Anyone reading the site closely this week likely noticed a new name entering the chat. We’re thrilled to welcome Rosie Thomas to the gang for the summer as an editorial intern! Rosie was previously a software engineer in the personal finance space. Currently halfway through her master’s degree in journalism,
     

Behind the Blog: The Omnipresence Is the Point

20 juin 2025 à 11:26
Behind the Blog: The Omnipresence Is the Point

This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss Deadheads and doxxing sites.

SAM: Anyone reading the site closely this week likely noticed a new name entering the chat. We’re thrilled to welcome Rosie Thomas to the gang for the summer as an editorial intern! 

Rosie was previously a software engineer in the personal finance space. Currently halfway through her master’s degree in journalism, Rosie is interested in social movements, how people change their behaviors in the face of new technologies, and “the infinite factors that influence sentiment and opinions,” in her words. In her program, she’s expanding her skills in investigations, audio production, and field recording. She published her first blog with us on day two, a really interesting (and in 404 style, informatively disturbing) breakdown of a new report that found tens of thousands of camera feeds exposed to the dark web. We’re so excited to see what she does with us this summer! 

  • ✇404 Media
  • ‘Martyrdom or Bust:’ Texas Man Caught Plotting Terror Attack Through Roblox Chats
    This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.The FBI has accused a Texas man, James Wesley Burger, of planning an Islamic State-style terrorist attack on a Christian music festival and talking about it on Roblox. The feds caught Burger after another Roblox user overheard his conversations about martyrdom and murder and tipped them off. The feds said that when they searched Burger’s phone they
     

‘Martyrdom or Bust:’ Texas Man Caught Plotting Terror Attack Through Roblox Chats

20 juin 2025 à 08:00
‘Martyrdom or Bust:’ Texas Man Caught Plotting Terror Attack Through Roblox Chats

This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.

The FBI has accused a Texas man, James Wesley Burger, of planning an Islamic State-style terrorist attack on a Christian music festival and talking about it on Roblox. The feds caught Burger after another Roblox user overheard his conversations about martyrdom and murder and tipped them off. The feds said that when they searched Burger’s phone they found a list of searches that included “ginger isis member” and “are suicide attacks haram in islam.” 

According to charging documents, a Roblox player contacted federal authorities after seeing another player called “Crazz3pain” talking about killing people. Screenshots from the server and included in the charging documents show Roblox avatars with beards dressed in Keffiyehs talking about dealing a “greivoius [sic] wound upon followers of the cross.”

“The witness observed the user of Crazz3pain state they were willing, as reported by the Witness, to ‘kill Shia Musilms at their mosque,” court records said. “Crazz3pain and another Roblox user[…]continued to make violent statements so the witness left the game.”

The witness stayed off of Roblox for two days and when they returned they saw Crazz3pain say something else that worried them, according to the court filing. “The Witness observed Crazz3pain tell Roblox User 1 to check their message on Discord,” the charging document said. “Roblox User 1 replied on Roblox to Crazz3pain, they should delete the photograph of firearms within the unknown Discord chat, ‘in case it was flagged as suspicious…the firearms should be kept hidden.”

According to the witness, Crazz3pain kept talking about their desire to commit “martyrdom” at a Christian event and that he wanted to “bring humiliation to worshippers of the cross.” The Witness allegedly asked Crazz3pain if the attack would happen at a church service and Crazz3pain told them it would happen at a concert. 

Someone asked Crazz3pain when it would happen. “‘It will be months…Shawwal…April,’” Crazz3pain said. Shawwal is the month after Ramadan in the Islamic calendar. The conversations the witness shared with the FBI happened on January 21 and 23, 2025.

Roblox gave authorities Crazz3pain’s email address, name, physical address, and IP address and it all pointed back to James Wesley Burger. The FBI searched Burger’s home on February 28 and discovered that someone in his family had put on a keylogger on the laptop he used to play Roblox and that they’d captured a lot of what he’d been typing while playing the game. They turned over the records to the feds.

“The safety of our community is among our highest priorities. In this case, we moved swiftly to assist law enforcement’s investigation before any real-world harm could occur and investigated and took action in accordance with our policies. We have a robust set of proactive and preventative safety measures designed to help swiftly detect and remove content that violates our policies," a spokesperson for Roblox told 404 Media. "Our Community Standards explicitly prohibit any content or behavior that depicts, supports, glorifies, or promotes terrorist or extremist organizations in any way. We have dedicated teams focused on proactively identifying and swiftly removing such content, as well as supporting requests from and providing assistance to law enforcement. We also work closely with other platforms and in close collaboration with safety organizations to keep content that violates our policies off our platform, and will continue to diligently enforce our policies.”

Burger’s plan to kill Christians was allegedly captured by the keylogger. “I’ve come to conclude it will befall the 12 of Shawwal aa/And it will be a music festival /Attracting bounties of Christians s/In’shaa’allah we will attain martyrdom /And deal a grevious [sic] wound upon the followers of the Cross /Pray for me and enjoin yourself to martyrdom,” he allegedly typed in Roblox, according to court records.

The FBI then interviewed Burger in his living room and he admitted he used the Crazz3pain account to play Roblox. The feds asked him about his alleged plan to kill Christians at a concert. Burger said it was, at the time, “mostly a heightened emotional response,” according to the court records. 

Burger also said that the details “became exaggerated” but that the goal “hasn’t shifted a bit,” according to the court records. He said he wanted to “[G]et the hell out of the U.S.” And if he can’t, “then, martyrdom or bust.”

He said that his intention with the attack “is something that is meant to or will cause terror,” according to the charging document. When the FBI agent asked if he was a terrorist, Burger said, “I mean, yeah, yeah. By, by the sense and … by my very own definition, yes, I guess, you know, I would be a terrorist.” 

When authorities searched his iPhone, they discovered two notes on the phone that described how to avoid leaving behind DNA and fingerprints at a crime scene. A third note appeared to be a note explaining the attack, meant to be read after it occured.

The list of previous searches on his iPhone included “Which month is april in islam,” “Festivals happening near me,” “are suicide attacks haram in islam,” “ginger isis member,” “lone wolf terrorists isis,” and “can tou kill a woman who foesnt[sic] wear hijab.”

Burger has been charged with making violent threats online and may spend time in a federal prison if convicted. This is not the first time something like this has happened on Roblox. The popular children’s game has been a popular spot for extremist behavior, including Nazis and religious terrorists, for years now. Last year, the DOJ accused a Syrian man living in Albanian of using Roblox to coordinate a group of American teenagers to disrupt public city council Zoom meetings.

  • ✇Dans les algorithmes
  • Ecrire le code du numérique
    C’est une formidable histoire que raconte le Code du numérique. Un livre édité par les Habitant.es des images ASBL et la Cellule pour la réduction des inégalités sociales et de la lutte contre la pauvreté de Bruxelles. Ce livre est le résultat de trois années d’action nées des difficultés qu’ont éprouvé les plus démunis à accéder à leurs droits durant la pandémie. En réaction à la fermeture des guichets d’aide sociale pendant la crise Covid, des militants du secteur social belge ont lancé un gro
     

Ecrire le code du numérique

19 juin 2025 à 01:00

C’est une formidable histoire que raconte le Code du numérique. Un livre édité par les Habitant.es des images ASBL et la Cellule pour la réduction des inégalités sociales et de la lutte contre la pauvreté de Bruxelles. Ce livre est le résultat de trois années d’action nées des difficultés qu’ont éprouvé les plus démunis à accéder à leurs droits durant la pandémie. En réaction à la fermeture des guichets d’aide sociale pendant la crise Covid, des militants du secteur social belge ont lancé un groupe de travail pour visibiliser le vécu collectif des souffrances individuelles des plus précaires face au déploiement du numérique, donnant naissance au Comité humain du numérique. “La digitalisation de la société n’a pas entraîné une amélioration généralisée des compétences numériques”, rappelle le Comité en s’appuyant sur le baromètre de l’inclusion numérique belge

Le Comité humain du numérique s’installe alors dans les quartiers et, avec les habitants, décide d’écrire un Code de loi : “Puisque l’Etat ne nous protège pas, écrivons les lois à sa place”. Rejoints par d’autres collectifs, le Comité humain se met à écrire la loi avec les habitants, depuis les témoignages de ceux qui n’arrivent pas à accomplir les démarches qu’on leur demande. Manifestations, séances d’écriture publique, délibérations publiques, parlement de rues… Le Comité implique les habitants, notamment contre l’ordonnance Bruxelles numérique qui veut rendre obligatoire les services publics digitalisés, sans garantir le maintien des guichets humains et rejoint la mobilisation coordonnée par le collectif Lire et écrire et plus de 200 associations. Devant le Parlement belge, le Comité humain organise des parlements humains de rue pour réclamer des guichets ! Suite à leur action, l’ordonnance Bruxelles numérique est amendée d’un nouvel article qui détermine des obligations pour les administrations à prévoir un accès par guichet, téléphone et voie postale – mais prévoit néanmoins la possibilité de s’en passer si les charges sont disproportionnées. Le collectif œuvre désormais à attaquer l’ordonnance devant la cour constitutionnelle belge et continue sa lutte pour refuser l’obligation au numérique.

Mais l’essentiel n’est pas que dans la victoire à venir, mais bien dans la force de la mobilisation et des propositions réalisées. Le Code du numérique ce sont d’abord 8 articles de lois amendés et discutés par des centaines d’habitants. L’article 1er rappelle que tous les services publics doivent proposer un accompagnement humain. Il rappelle que “si un robot ne nous comprend pas, ce n’est pas nous le problème”. Que cet accès doit être sans condition, c’est-à-dire gratuit, avec des temps d’attente limités, “sans rendez-vous”, sans obligation de maîtrise de la langue ou de l’écriture. Que l’accompagnement humain est un droit. Que ce coût ne doit pas reposer sur d’autres, que ce soit les proches, les enfants, les aidants ou les travailleurs sociaux. Que l’Etat doit veiller à cette accessibilité humaine et qu’il doit proposer aux citoyen.nes des procédures gratuites pour faire valoir leurs droits. L’article 2 rappelle que c’est à l’Etat d’évaluer l’utilité et l’efficacité des nouveaux outils numériques qu’il met en place : qu’ils doivent aider les citoyens et pas seulement les contrôler. Que cette évaluation doit associer les utilisateurs, que leurs impacts doivent être contrôlés, limités et non centralisés. L’article 3 rappelle que l’Etat doit créer ses propres outils et que les démarches administratives ne peuvent pas impliquer le recours à un service privé. L’article 4 suggère de bâtir des alternatives aux solutions numériques qu’on nous impose. L’article 5 suggère que leur utilisation doit être contrainte et restreinte, notamment selon les lieux ou les âges et souligne que l’apprentissage comme l’interaction entre parents et écoles ne peut être conditionnée par des outils numériques. L’article 6 en appelle à la création d’un label rendant visible le niveau de dangerosité physique ou mentale des outils, avec des possibilités de signalement simples. L’article 7 milite pour un droit à pouvoir se déconnecter sans se justifier. Enfin, l’article 8 plaide pour une protection des compétences humaines et de la rencontre physique, notamment dans le cadre de l’accès aux soins. “Tout employé.e/étudiant.e/patient.e/client.e a le droit d’exiger de rencontrer en face à face un responsable sur un lieu physique”. L’introduction de nouveaux outils numériques doit être développée et validée par ceux qui devront l’utiliser.

Derrière ces propositions de lois, simples, essentielles… la vraie richesse du travail du Comité humain du numérique est de proposer, de donner à lire un recueil de paroles qu’on n’entend nulle part. Les propos des habitants, des individus confrontés à la transformation numérique du monde, permettent de faire entendre des voix qui ne parviennent plus aux oreilles des concepteurs du monde. Des paroles simples et fortes. Georges : “Ce que je demanderai aux politiciens ? C’est de nous protéger de tout ça.” Anthony : “Internet devait être une plateforme et pas une vie secondaire”. Nora : “En tant qu’assistante sociale, le numérique me surresponsabilise et rend le public surdépendant de moi. Je suis le dernier maillon de la chaîne, l’échec social passe par moi. Je le matérialise”. Amina : “Je ne sais pas lire, je ne sais pas écrire. Mais je sais parler. Le numérique ne me laisse pas parler”. Aïssatou : “Maintenant tout est trop difficile. S’entraider c’est la vie. Avec le numérique il n’y a plus personne pour aider”. Khalid : “Qu’est-ce qui se passe pour les personnes qui n’ont pas d’enfant pour les aider ?” Elise : “Comment s’assurer qu’il n’y a pas de discrimination ?” Roger : “Le numérique est utilisé pour décourager les démarches”, puisque bien souvent on ne peut même pas répondre à un courriel. AnaÎs : “Il y a plein d’infos qui ne sont pas numérisées, car elles n’entrent pas dans les cases. La passation d’information est devenue très difficile”… Le Code du numérique nous “redonne à entendre les discours provenant des classes populaires”, comme nous y invitait le chercheur David Gaborieau dans le rapport “IA : la voie citoyenne”.

Le Code du numérique nous rappelle que désormais, les institutions s’invitent chez nous, dans nos salons, dans nos lits. Il rappelle que l’accompagnement humain sera toujours nécessaire pour presque la moitié de la population. Que “l’aide au remplissage” des documents administratifs ne peut pas s’arrêter derrière un téléphone qui sonne dans le vide. Que “la digitalisation des services publics et privés donne encore plus de pouvoir aux institutions face aux individus”. Que beaucoup de situations n’entreront jamais dans les “cases” prédéfinies.Le Code du numérique n’est pas qu’une expérience spécifique et située, rappellent ses porteurs. “Il est là pour que vous vous en empariez”. Les lois proposées sont faites pour être débattues, modifiées, amendées, adaptées. Les auteurs ont créé un jeu de cartes pour permettre à d’autres d’organiser un Parlement humain du numérique. Il détaille également comment créer son propre Comité humain, invite à écrire ses propres lois depuis le recueil de témoignages des usagers, en ouvrant le débat, en écrivant soi-même son Code, ses lois, à organiser son parlement et documente nombre de méthodes et d’outils pour interpeller, mobiliser, intégrer les contributions. Bref, il invite à ce que bien d’autres Code du numérique essaiment, en Belgique et bien au-delà ! A chacun de s’en emparer.

Cet article a été publié originellement pour la lettre d’information du Conseil national du numérique du 23 mai 2025.

Le Code du numérique.
  • ✇404 Media
  • One of the Universe’s Biggest Mysteries Has Been Solved, Scientists Say
    🌘Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week. Scientists have directly confirmed the location of the universe's “missing” matter for the first time, reports a study published on Monday in Nature Astronomy. The idea that the universe must contain normal, or “baryonic,” matter that we can’t seem to find goes back to the birth of modern cosmological models. Now, a team has revealed that about 76 percent of
     

One of the Universe’s Biggest Mysteries Has Been Solved, Scientists Say

18 juin 2025 à 09:47
🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
One of the Universe’s Biggest Mysteries Has Been Solved, Scientists Say

Scientists have directly confirmed the location of the universe's “missing” matter for the first time, reports a study published on Monday in Nature Astronomy

The idea that the universe must contain normal, or “baryonic,” matter that we can’t seem to find goes back to the birth of modern cosmological models. Now, a team has revealed that about 76 percent of all baryons—the ordinary particles that make up planets and stars—exist as gas hidden in the dark expanses between galaxies, known as the intergalactic medium. Fast radio bursts (FRBs), transient signals with elusive origins, illuminated the missing baryons, according to the researchers. As a bonus, they also identified the most distant FRB ever recorded, at 9.1 billion light years away, in the study. 

“Measuring the ‘missing baryons’ with Fast Radio Bursts has been a major long-sought milestone for radio astronomers,” said Liam Connor, an astronomer at the Center for Astrophysics | Harvard & Smithsonian who led the study, in an email. “Until recently, we didn’t have a large-enough sample of bursts to make strong statements about where this ordinary matter was hiding.” 

Under the leadership of Caltech professor Vikram Ravi, the researchers constructed the DSA-110 radio telescope—an array of over 100 dishes in the California desert—to achieve this longstanding milestone. “We built up the largest and most distant collection of localized FRBs (meaning we know their exact host galaxy and distance),” Connor explained. “This data sample, plus new algorithms, allowed us to finally make a complete baryon pie chart. There are no longer any missing wedges.”

Baryons are the building blocks of the familiar matter that makes up our bodies, stars, and galaxies, in contrast to dark matter, a mysterious substance that accounts for the vast majority of the universe’s mass. Cosmological models predict that there is much more baryonic matter than we can see in stars and galaxies, which has spurred astronomers into a decades-long search for the “missing baryons” in space. 

Scientists have long assumed that most of this missing matter exists in the form of ionized gas in the IGM, but FRBs have opened a new window into these dark reaches, which can be difficult to explore with conventional observatories. 

“FRBs complement and improve on past methods by their sensitivity to all the ionized gas in the Universe,” Connor said. “Past methods, which were highly informative but somewhat incomplete, could only measure hot gas near galaxies or clusters of galaxies. There was no probe that could measure the lion’s share of ordinary matter in the Universe, which it turns out is in the intergalactic medium.”

Since the first FRB was detected in 2007, thousands of similar events have been discovered, though astronomers still aren't sure what causes them. Characterized by extremely energetic radio waves that last for mere milliseconds, the bursts typically originate millions or billions of light years from our galaxy. Some repeat, and some do not. Scientists think these pyrotechnic events are fueled by massive compact objects, like neutron stars, but their exact nature and origins remain unclear.

Connor and his colleagues studied a sample of 60 FRB observations that spanned from about 12 million light years away from Earth all the way to a new record holder for distance: FRB 20230521B, located 9.1 billion light years away. With the help of these cosmic searchlights, the team was able to make a new precise measurement of the density of baryonic matter across the cosmic web, which is a network of large-scale structures that spans the universe. The results matched up with cosmological predictions that most of the missing baryons would be blown out into the IGM by “feedback” generated within galaxies. About 15 percent is present in structures that surround galaxies, called halos, and a small remainder makes up stars and other celestial bodies.

 

“It really felt like I was going in blind without a strong prior either way,” Connor said. “If all of the missing baryons were hiding in galaxy halos and the IGM were gas-poor, that would be surprising in its own way. If, as we discovered, the baryons had mostly been blown into the space between galaxies, that would also be remarkable because that would require strong astrophysical feedback and violent processes during galaxy formation.”

“Now, looking back on the result, it’s kind of satisfying that our data agrees with modern cosmological simulations with strong ‘feedback’ and agrees with the early Universe values of the total abundance of normal matter,” he continued. “Sometimes it’s nice to have some concordance.” 

The new measurement might alleviate the so-called sigma-8 tension, which is a discrepancy between the overall “clumpiness” of matter in the universe when measured using the cosmic microwave background, which is the oldest light in the cosmos, compared with using modern maps of galaxies and clusters.

“One explanation for this disagreement is that our standard model of cosmology is broken, and we need exotic new physics,” Connor said. “Another explanation is that today’s Universe appears smooth because the baryons have been sloshed around by feedback.” 

“Our FRB measurement suggests the baryon cosmic web is relatively smooth, homogenized by astrophysical processes in galaxies (feedback),” he continued. “This would explain the S8 tension without exotic new physics. If that’s the case, then I think the broader lesson is that we really need to pin down these pesky baryons, which have previously been very difficult to measure directly.”

To that end, Connor is optimistic that more answers to these cosmic riddles are coming down the pike. 

“The future is looking bright for the field of FRB cosmology,” he said. “We are in the process of building enormous radio telescope arrays that could find tens of thousands of localized FRBs each year,” including the upcoming DSA-2000.

“My colleagues and I think of our work as baby steps towards the bigger goal of fully mapping the ordinary, baryonic matter throughout the whole Universe,” he concluded. 

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.

  • ✇404 Media
  • 40,000 Cameras, From Bird Feeders to Baby Monitors, Exposed to the Internet
    A report from a cybersecurity company last week found that over 40,000 unsecured cameras—including CCTV and security cameras on public transportation, in hospitals, on internet-connected bird feeders and on ATMs—are exposed online worldwide. Cybersecurity risk intelligence company BitSight was able to access and download content from thousands of internet-connected systems, including domestic and commercial webcams, baby monitors, office security, and pet cams. They also found content from th
     

40,000 Cameras, From Bird Feeders to Baby Monitors, Exposed to the Internet

18 juin 2025 à 09:40
40,000 Cameras, From Bird Feeders to Baby Monitors, Exposed to the Internet

A report from a cybersecurity company last week found that over 40,000 unsecured cameras—including CCTV and security cameras on public transportation, in hospitals, on internet-connected bird feeders and on ATMs—are exposed online worldwide. 

Cybersecurity risk intelligence company BitSight was able to access and download content from thousands of internet-connected systems, including domestic and commercial webcams, baby monitors, office security, and pet cams. They also found content from these cameras on locations on the dark web where people share and sell access to their live feeds. “The most concerning examples found were cameras in hospitals or clinics monitoring patients, posing a significant privacy risk due to the highly sensitive nature of the footage,” said João Cruz, Principal Security Research Scientist for the team that produced the report.

The company wrote in a press release that it “doesn’t take elite hacking to access these cameras; in most cases, a regular web browser and a curious mind are all it takes, meaning that 40,000 figure is probably just the tip of the iceberg.” 

Depending on the type of login protocol that the cameras were using, the researchers were able to access footage or individual real-time screenshots. Against a background of increasing surveillance by law enforcement and ICE, there is clear potential for abuse of unknowingly open cameras. 

Traffic Camera ‘Selfie’ Creator Holds Cease and Desist Letter in Front of Traffic Cam
Traffic Cam Photobooth lets you take a capture from NYC surveillance camera. The city’s Department of Transportation does not like that.
40,000 Cameras, From Bird Feeders to Baby Monitors, Exposed to the Internet404 MediaSamantha Cole
40,000 Cameras, From Bird Feeders to Baby Monitors, Exposed to the Internet

“Knowing the real number is practically impossible due to the insanely high number of camera brands and models existent in the market,” said Cruz, “each of them with different ways to check if it’s exposed and if it’s possible to get access to the live footage.”

The report outlines more obvious risks, from tracking the behavioral patterns and real-time status of when people are in their homes in order to plan a burglary, to “shoulder surfing,” or stealing data by observing someone logging in to a computer in offices. The report also found cameras in stores, gyms, laundromats, and construction sites, meaning that exposed cameras are monitoring people in their daily lives. The geographic data provided by the camera’s IP addresses, combined with commercially available facial-recognition systems, could prove dangerous for individuals working in or using those businesses.

You can find out if your camera has been exposed using a site like Shodan.io, a search engine which scans for devices connected to the internet, or by trying to access your camera from a device logged in to a different network. Users should also check the documentation provided by the manufacturer, rather than just plugging in a camera right away, to minimize vulnerabilities, and make sure that they set their own password on any IoT-connected device. 

This is because many brands use default logins for their products, and these logins are easily findable online. The BitSight report didn’t try to hack into these kinds of cameras, or try to brute-force any passwords, but, “if we did so, we firmly believe that the number would be higher,” said Cruz. Older camera systems with deprecated and unmaintained software are more susceptible to being hacked in this way; one somewhat brighter spot is that these “digital ghost ships” seem to be decreasing in number as the oldest and least secure among them are replaced or fail completely. 

Unsecured cameras attract hackers and malicious actors, and the risks can go beyond the embarrassing, personal, or even individual. In March this year, the hacking group Akira successfully compromised an organisation using an unsecured webcam, after a first attack attempt was effectively prevented by cybersecurity protocols. In 2024, the Ukrainian government asked citizens to turn off all broadcasting cameras, after Russian agents hacked into webcams at a condo association and a car park. They altered the direction of the cameras to point toward nearby infrastructure and used the footage in planning strikes. Ukraine blocked the operation of 10,000 internet-connected digital security cameras in order to prevent further information leaks, and a May 2025 report from the Joint Cybersecurity Advisory described continued attacks from Russian espionage units on private and municipal cameras to track materials entering Ukraine.

  • ✇404 Media
  • The AI Slop Fight Between Iran and Israel
    As Israel and Iran trade blows in a quickly escalating conflict that risks engulfing the rest of the region as well as a more direct confrontation between Iran and the U.S., social media is being flooded with AI-generated media that claims to show the devastation, but is fake.The fake videos and images show how generative AI has already become a staple of modern conflict. On one end, AI-generated content of unknown origin is filling the void created by state-sanctioned media blackouts with misin
     

The AI Slop Fight Between Iran and Israel

18 juin 2025 à 09:00
The AI Slop Fight Between Iran and Israel

As Israel and Iran trade blows in a quickly escalating conflict that risks engulfing the rest of the region as well as a more direct confrontation between Iran and the U.S., social media is being flooded with AI-generated media that claims to show the devastation, but is fake.

The fake videos and images show how generative AI has already become a staple of modern conflict. On one end, AI-generated content of unknown origin is filling the void created by state-sanctioned media blackouts with misinformation, and on the other end, the leaders of these countries are sharing AI-generated slop to spread the oldest forms of xenophobia and propaganda.

If you want to follow a war as it’s happening, it’s easier than ever. Telegram channels post live streams of bombing raids as they happen and much of the footage trickles up to X, TikTok, and other social media platforms. There’s more footage of conflict than there’s ever been, but a lot of it is fake.

A few days ago, Iranian news outlets reported that Iran’s military had shot down three F-35s. Israel denied it happened. As the claim spread so did supposed images of the downed jet. In one, a massive version of the jet smolders on the ground next to a town. The cockpit dwarfs the nearby buildings and tiny people mill around the downed jet like Lilliputians surrounding Gulliver.

It’s a fake, an obvious one, but thousands of people shared it online. Another image of the supposedly downed jet showed it crashed in a field somewhere in the middle of the night. Its wings were gone and its afterburner still glowed hot. This was also a fake.

The AI Slop Fight Between Iran and Israel
Image via X.com.
The AI Slop Fight Between Iran and Israel
Image via X.com.

AI slop is not the sole domain of anonymous amateur and professional propagandists. The leaders of both Iran and Israel are doing it too. The Supreme Leader of Iran is posting AI-generated missile launches on his X account, a match for similar grotesques on the account of Israel’s Minister of Defense.

New tools like Google’s Veo 3 make AI-generated videos more realistic than ever. Iranian news outlet Tehran Times shared a video to X that it said captured “the moment an Iranian missile hit a building in Bat Yam, southern Tel Aviv.” The video was fake. In another that appeared to come from a TV news spot, a massive missile moved down a long concrete hallway. It’s also clearly AI-generated, and still shows the watermark in the bottom right corner for Veo.

#BREAKING
Doomsday in Tel Aviv pic.twitter.com/5CDSUDcTY0

— Tehran Times (@TehranTimes79) June 14, 2025

After Iran launched a strike on Israel, Tehran Times shared footage of what it claimed was “Doomsday in Tel Aviv.” A drone shot rotated through scenes of destroyed buildings and piles of rubble. Like the other videos, it was an AI generated fake that appeared on both a Telegram account and TikTok channel named “3amelyonn.”

In Arabic, 3amelyonn’s TikTok channel calls itself “Artificial Intelligence Resistance” but has no such label on Telegram. It’s been posting on Telegram since 2023 and its first TikTok video appeared in April of 2025, of an AI-generated tour through Lebanon, showing its various cities as smoking ruins. It’s full of the quivering lines and other hallucinations typical of early AI video.

But 3amelyonn’s videos a month later are more convincing. A video posted on June 5, labeled as Ben Gurion Airport, shows bombed out buildings and destroyed airplanes. It’s been viewed more than 2 million times. The video of a destroyed Tel Aviv, the one that made it on to Tehran Times, has been viewed more than 11 million times and was posted on May 27, weeks before the current conflict.

Hany Farid, a UC Berkeley professor and founder of GetReal, a synthetic media detection company, has been collecting these fake videos and debunking them. 

“In just the last 12 hours, we at GetReal have been seeing a slew of fake videos surrounding the recent conflict between Israel and Iran. We have been able to link each of these visually compelling videos to Veo 3,” he said in a post on LinkedIn. “It is no surprise that as generative-AI tools continue to improve in photo-realism, they are being misused to spread misinformation and sow confusion.”

The spread of AI-generated media about this conflict appears to be particularly bad because both Iran and Israel are asking their citizens not to share media of destruction, which may help the other side with its targeting for future attacks. On Saturday, for example, the Israel Defense Force asked people not to “publish and share the location or documentation of strikes. The enemy follows these documentations in order to improve its targeting abilities. Be responsible—do not share locations on the web!” Users on social media then fill this vacuum with AI-generated media.

“The casualty in this AI war [is] the truth,” Farid told 404 Media. “By muddying the waters with AI slop, any side can now claim that any other videos showing, for example, a successful strike or human rights violations are fake. Finding the truth at times of conflict has always been difficult, and now in the age of AI and social media, it is even more difficult.”

“We're committed to developing AI responsibly and we have clear policies to protect users from harm and governing the use of our AI tools,” a Google spokesperson told 404 Media. “Any content generated with Google AI has a SynthID watermark embedded and we add a visible watermark to Veo videos too.”

Farid and his team used SynthID to identify the fake videos “alongside other forensic techniques that we have developed over at GetReal,” he said. But checking a video for a SynthID watermark, which is visually imperceptible, requires someone to take the time to download the video and upload it to a separate website. Casual social media scrollers are not taking the time to verify a video they’re seeing by sending it to the SynthID website.

One distinguishing feature of 3amelyonn and others’ videos of viral AI slop about the conflict is that the destruction is confined to buildings. There are no humans and no blood in 3amelyonn’s  aerial shots of destruction, which are more likely to get blocked both by AI image and video generators as well as the social media platforms where these creations are shared. If a human does appear, they’re as observers like in the F-35 picture or milling soldiers like the tunnel video. Seeing a soldier in active combat or a wounded person is rare.

There’s no shortage of real, horrifying footage from Gaza and other conflicts around the world. AI war spam, however, is almost always bloodless. A year ago, the AI-generated image “All Eyes on Raffah” garnered tens of millions of views. It was created by a Facebook group with the goal of “Making AI prosper.”

  • ✇404 Media
  • Podcast: Airlines Sold Your Flight Data to DHS—And Covered It Up
    This week we start with Joseph’s article about the U.S’s major airlines selling customers’ flight information to Customs and Border Protection and then telling the agency to not reveal where the data came from. After the break, Emanuel tells us how AI scraping bots are breaking open libraries, archives, and museums. In the subscribers-only section, Jason explains the casual surveillance relationship between ICE and local cops, according to emails he got. Listen to the weekly podcast on App
     

Podcast: Airlines Sold Your Flight Data to DHS—And Covered It Up

18 juin 2025 à 09:00
Podcast: Airlines Sold Your Flight Data to DHS—And Covered It Up

This week we start with Joseph’s article about the U.S’s major airlines selling customers’ flight information to Customs and Border Protection and then telling the agency to not reveal where the data came from. After the break, Emanuel tells us how AI scraping bots are breaking open libraries, archives, and museums. In the subscribers-only section, Jason explains the casual surveillance relationship between ICE and local cops, according to emails he got.

Listen to the weekly podcast on Apple Podcasts, Spotify, or YouTube. Become a paid subscriber for access to this episode's bonus content and to power our journalism. If you become a paid subscriber, check your inbox for an email from our podcast host Transistor for a link to the subscribers-only version! You can also add that subscribers feed to your podcast app of choice and never miss an episode that way. The email should also contain the subscribers-only unlisted YouTube link for the extended video version too. It will also be in the show notes in your podcast player.

  • ✇Dans les algorithmes
  • Chatbots, une adoption sans impact ?
    Dans sa dernière newsletter, Algorithm Watch revient sur une étude danoise qui a observé les effets des chatbots sur le travail auprès de 25 000 travailleurs provenant de 11 professions différentes où des chatbots sont couramment utilisés (développeurs, journalistes, professionnels RH, enseignants…). Si ces travailleurs ont noté que travailler avec les chatbots leur permettait de gagner du temps, d’améliorer la qualité de leur travail, le gain de temps s’est avéré modeste, représentant seulement
     

Chatbots, une adoption sans impact ?

18 juin 2025 à 01:00

Dans sa dernière newsletter, Algorithm Watch revient sur une étude danoise qui a observé les effets des chatbots sur le travail auprès de 25 000 travailleurs provenant de 11 professions différentes où des chatbots sont couramment utilisés (développeurs, journalistes, professionnels RH, enseignants…). Si ces travailleurs ont noté que travailler avec les chatbots leur permettait de gagner du temps, d’améliorer la qualité de leur travail, le gain de temps s’est avéré modeste, représentant seulement 2,8% du total des heures de travail. La question des gains de productivité de l’IA générative dépend pour l’instant beaucoup des études réalisées, des tâches et des outils. Les gains de temps varient certes un peu selon les profils de postes (plus élevés pour les professions du marketing (6,8%) que pour les enseignants (0,2%)), mais ils restent bien modestes.”Sans flux de travail modifiés ni incitations supplémentaires, la plupart des effets positifs sont vains”

Algorithm Watch se demande si les chatbots ne sont pas des outils de travail improductifs. Il semblerait plutôt que, comme toute transformation, elle nécessite surtout des adaptations organisationnelles ad hoc pour en développer les effets.

  • ✇404 Media
  • California Cops Investigate ‘Immigration Protest’ With AI-Camera System
    📄This article was primarily reported using public records requests. We are making it available to all readers as a public service. FOIA reporting can be expensive, please consider subscribing to 404 Media to support this work. Or send us a one time donation via our tip jar here.A California police department searched AI-enabled, automatic license plate reader (ALPR) cameras in relation to an “immigration protest,” according to internal police data obtained by 404 Media. The data also shows that
     

California Cops Investigate ‘Immigration Protest’ With AI-Camera System

17 juin 2025 à 11:15
📄
This article was primarily reported using public records requests. We are making it available to all readers as a public service. FOIA reporting can be expensive, please consider subscribing to 404 Media to support this work. Or send us a one time donation via our tip jar here.
California Cops Investigate ‘Immigration Protest’ With AI-Camera System

A California police department searched AI-enabled, automatic license plate reader (ALPR) cameras in relation to an “immigration protest,” according to internal police data obtained by 404 Media. The data also shows that police departments and sheriff offices around the country have repeatedly tapped into the cameras inside California, made by a company called Flock, on behalf of Immigration and Customs Enforcement (ICE), digitally reaching into the sanctuary state in a data sharing practice that experts say is illegal. 

Flock allows participating agencies to search not only cameras in their jurisdiction or state, but nationwide, meaning that local police that may work directly with ICE on immigration enforcement are able to search cameras inside California or other states. But this data sharing is only possible because California agencies have opted-in to sharing it with agencies in other states, making them legally responsible for the data sharing. 

The news raises questions about whether California agencies are enforcing the law on their own data sharing practices, threatens to undermine the state’s perception as a sanctuary state, and highlights the sort of surveillance or investigative tools law enforcement may deploy at immigration related protests. Over the weekend, millions of people attended No Kings protests across the U.S. 404 Media’s findings come after we revealed police were searching cameras in Illinois on behalf of ICE, and then Cal Matters found local law enforcement agencies in California were searching cameras for ICE too.

Meta Invents New Way to Humiliate Users With Feed of People's Chats With AI

17 juin 2025 à 10:47
Subscribe
Join the newsletter to get the latest updates.
Success
Great! Check your inbox and click the link.
Error
Please enter a valid email address.
Meta Invents New Way to Humiliate Users With Feed of People's Chats With AI

I was sick last week, so I did not have time to write about the Discover Tab in Meta’s AI app, which, as Katie Notopoulos of Business Insider has pointed out, is the “saddest place on the internet.” Many very good articles have already been written about it, and yet, I cannot allow its existence to go unremarked upon in the pages of 404 Media. 

If you somehow missed this while millions of people were protesting in the streets, state politicians were being assassinated, war was breaking out between Israel and Iran, the military was deployed to the streets of Los Angeles, and a Coinbase-sponsored military parade rolled past dozens of passersby in Washington, D.C., here is what the “Discover” tab is: The Meta AI app, which is the company’s competitor to the ChatGPT app, is posting users’ conversations on a public “Discover” page where anyone can see the things that users are asking Meta’s chatbot to make for them. 

Meta Invents New Way to Humiliate Users With Feed of People's Chats With AI

This includes various innocuous image and video generations that have become completely inescapable on all of Meta’s platforms (things like “egg with one eye made of black and gold,” “adorable Maltese dog becomes a heroic lifeguard,” “one second for God to step into your mind”), but it also includes entire chatbot conversations where users are seemingly unknowingly leaking a mix of embarrassing, personal, and sensitive details about their lives onto a public platform owned by Mark Zuckerberg. In almost all cases, I was able to trivially tie these chats to actual, real people because the app uses your Instagram or Facebook account as your login.

  • ✇404 Media
  • The People Search Sites in the Suspected Minnesota Killer's Notebook Are a Failure of Congress
    On Monday, federal and state authorities charged Vance Boelter with the murders of Minnesota Rep. Melissa Hortman and her husband. An affidavit written by an FBI Special Agent, published here by MSNBC, includes photos of a notepad found in Boelter’s SUV which included a long list of people search sites, some of which make it very easy for essentially anyone to find the address and other personal information of someone else in the U.S. The SUV contained other notebooks and some pages included
     

The People Search Sites in the Suspected Minnesota Killer's Notebook Are a Failure of Congress

17 juin 2025 à 10:23
The People Search Sites in the Suspected Minnesota Killer's Notebook Are a Failure of Congress

On Monday, federal and state authorities charged Vance Boelter with the murders of Minnesota Rep. Melissa Hortman and her husband. An affidavit written by an FBI Special Agent, published here by MSNBC, includes photos of a notepad found in Boelter’s SUV which included a long list of people search sites, some of which make it very easy for essentially anyone to find the address and other personal information of someone else in the U.S. The SUV contained other notebooks and some pages included the names of more than 45 Minnesota state and federal public officials, including Hortman, the affidavit says. Hortman’s home address was listed next to her name, it adds.

People search sites can present a risk to citizen’s privacy, and, depending on the context, physical safety. They aggregate data from property records, social media, marriage licenses, and other places and make it accessible to even those with no tech savvy. Some are free, some are paid, and some require a user to tick a box confirming they’re only using the data for certain permitted use cases. 

Congress has known about the risk of data for decades. In 1994 lawmakers created the Driver’s Privacy Protection Act (DPPA) after a stalker hired a private investigator who then obtained the address of actress Rebecca Schaeffer from a DMV. The stalker then murdered Schaeffer. With people search sites, though, lawmakers have been largely motionless, despite them existing for years, on the open web, accessible by a Google search and sometimes even promoted with Google advertisements.

Senator Ron Wyden said in a statement “The accused Minneapolis assassin allegedly used data brokers as a key part of his plot to track down and murder Democratic lawmakers. Congress doesn't need any more proof that people are being killed based on data for sale to anyone with a credit card. Every single American's safety is at risk until Congress cracks down on this sleazy industry.”

This notepad does not necessarily mean that Boelter used these specific sites to find Hortman’s or other officials’ addresses. As the New York Times noted, Hortman’s address was on her campaign website, and Minnesota State Senator John Hoffman, who Boelter allegedly shot along with Hoffman’s wife, listed his address on his official legislative webpage.

The sites’ inclusion shows they are of high interest to a person who allegedly murdered and targeted multiple officials and their families in an act of political violence. Next to some of the people search site names, Boelter appears to have put a star or tick.

Those people search sites are:

A spokesperson for Atlas, a company that is suing a variety of people search sites, said “Tragedies like this might be prevented if data brokers simply complied with state and federal privacy laws. Our company has been in court for more than 15 months litigating against each of the eleven data brokers identified in the alleged shooter’s writings, seeking to hold them accountable for refusing to comply with New Jersey’s Daniel’s Law which seeks to protect the home addresses of judges, prosecutors, law enforcement and their families. This industry’s purposeful refusal to comply with privacy laws has and continues to endanger thousands of public servants and their families.” 

404 Media has repeatedly reported on how data can be weaponized against people. We found violent criminals and hackers were able to dox nearly anyone in the U.S. for $15, using bots that were based on data people had given as part of opening credit cards. In 2023 Verizon gave sensitive information, including an address on file, of one of its customers to her stalker, who then drove to the address armed with a knife.

404 Media was able to contact most of the people search sites for comment. None responded.

Update: this piece has been updated to include a statement from Atlas. An earlier version of this piece accidentally published a version with a different structure; this correct version includes more information about the DPPA.

  • ✇404 Media
  • AI Scraping Bots Are Breaking Open Libraries, Archives, and Museums
    AI bots that scrape the internet for training data are hammering the servers of libraries, archives, museums, and galleries, and are in some cases knocking their collections offline, according to a new survey published today. While the impact of AI bots on open collections has been reported anecdotally, the survey is the first attempt at measuring the problem, which in the worst cases can make valuable, public resources unavailable to humans because the servers they’re hosted on are being swampe
     

AI Scraping Bots Are Breaking Open Libraries, Archives, and Museums

17 juin 2025 à 06:00
AI Scraping Bots Are Breaking Open Libraries, Archives, and Museums

AI bots that scrape the internet for training data are hammering the servers of libraries, archives, museums, and galleries, and are in some cases knocking their collections offline, according to a new survey published today. While the impact of AI bots on open collections has been reported anecdotally, the survey is the first attempt at measuring the problem, which in the worst cases can make valuable, public resources unavailable to humans because the servers they’re hosted on are being swamped by bots scraping the internet for AI training data. 

  • ✇Dans les algorithmes
  • Pour une science de la subjectivité
    « J’aimerais vous confronter à un problème de calcul difficile », attaque Albert Moukheiber sur la scène de la conférence USI 2025. « Dans les sciences cognitives, on est confronté à un problème qu’on n’arrive pas à résoudre : la subjectivité ! »  Le docteur en neuroscience et psychologue clinicien, auteur de Votre cerveau vous joue des tours (Allary éditions 2019) et de Neuromania (Allary éditions, 2024), commence par faire un rapide historique de ce qu’on sait sur le cerveau.  Où est le
     

Pour une science de la subjectivité

17 juin 2025 à 01:00

« J’aimerais vous confronter à un problème de calcul difficile », attaque Albert Moukheiber sur la scène de la conférence USI 2025. « Dans les sciences cognitives, on est confronté à un problème qu’on n’arrive pas à résoudre : la subjectivité ! » 

Le docteur en neuroscience et psychologue clinicien, auteur de Votre cerveau vous joue des tours (Allary éditions 2019) et de Neuromania (Allary éditions, 2024), commence par faire un rapide historique de ce qu’on sait sur le cerveau. 

Où est le neurone ?

« Contrairement à d’autres organes, un cerveau mort n’a rien à dire sur son fonctionnement. Et pendant très longtemps, nous n’avons pas eu d’instruments pour comprendre un cerveau ». En fait, les technologies permettant d’ausculter le cerveau, de cartographier son activité, sont assez récentes et demeurent bien peu précises. Pour cela, il faut être capable de mesurer son activité, de voir où se font les afflux d’énergie et l’activité chimique. C’est seulement assez récemment, depuis les années 1990 surtout, qu’on a développé des technologies pour étudier cette activité, avec les électro-encéphalogrammes, puis avec l’imagerie par résonance magnétique (IRM) structurelle et surtout fonctionnelle. L’IRM fonctionnelle est celle que les médecins vous prescrivent. Elle mesure la matière cérébrale permettant de créer une image en noir et blanc pour identifier des maladies, des lésions, des tumeurs. Mais elle ne dit rien de l’activité neuronale. Seule l’IRM fonctionnelle observe l’activité, mais il faut comprendre que les images que nous en produisons sont peu précises et demeurent probabilistes. Les images de l’IRMf font apparaître des couleurs sur des zones en activité, mais ces couleurs ne désignent pas nécessairement une activité forte de ces zones, ni que le reste du cerveau est inactif. L’IRMf tente de montrer que certaines zones sont plus actives que d’autres parce qu’elles sont plus alimentées en oxygène et en sang. L’IRMf fonctionne par soustraction des images passées. Le patient dont on mesure l’activité cérébrale est invité à faire une tâche en limitant au maximum toute autre activité que celle demandée et les scientifiques comparent  ces images à des précédentes pour déterminer quelles zones sont affectées quand vous fermez le poing par exemple. « On applique des calculs de probabilité aux soustractions pour tenter d’isoler un signal dans un océan de bruits », précise Moukheiber dans Neuromania. L’IRMf n’est donc pas un enregistrement direct de l’activation cérébrale pour une tâche donnée, mais « une reconstruction a posteriori de la probabilité qu’une aire soit impliquée dans cette tâche ». En fait, les couleurs indiquent des probabilités. « Ces couleurs n’indiquent donc pas une intensité d’activité, mais une probabilité d’implication ». Enfin, les mesures que nous réalisons n’ont rien de précis, rappelle le chercheur. La précision de l’IRMf est le voxel, qui contient environ 5,5 millions de neurones ! Ensuite, l’IRMf capture le taux d’oxygène, alors que la circulation sanguine est bien plus lente que les échanges chimiques de nos neurones. Enfin, le traitement de données est particulièrement complexe. Une étude a chargé plusieurs équipes d’analyser un même ensemble de données d’IRMf et n’a pas conduit aux mêmes résultats selon les équipes. Bref, pour le dire simplement, le neurone est l’unité de base de compréhension de notre cerveau, mais nos outils ne nous permettent pas de le mesurer. Il faut dire qu’il n’est pas non plus le bon niveau explicatif. Les explications établies à partir d’images issues de l’IRMf nous donnent donc plus une illusion de connaissance réelle qu’autre chose. D’où l’enjeu à prendre les résultats de nombre d’études qui s’appuient sur ces images avec beaucoup de recul. « On peut faire dire beaucoup de choses à l’imagerie cérébrale » et c’est assurément ce qui explique qu’elle soit si utilisée.

Les données ne suffisent pas

Dans les années 50-60, le courant de la cybernétique pensait le cerveau comme un organe de traitement de l’information, qu’on devrait étudier comme d’autres machines. C’est la naissance de la neuroscience computationnelle qui tente de modéliser le cerveau à l’image des machines. Outre les travaux de John von Neumann, Claude Shannon prolonge ces idées d’une théorie de l’information qui va permettre de créer des « neurones artificiels », qui ne portent ce nom que parce qu’ils ont été créés pour fonctionner sur le modèle d’un neurone. En 1957, le Perceptron de Frank Rosenblatt est considéré comme la première machine à utiliser un réseau neuronal artificiel. Mais on a bien plus appliqué le lexique du cerveau aux ordinateurs qu’autre chose, rappelle Albert Moukheiber. 

Aujourd’hui, l’Intelligence artificielle et ses « réseaux de neurones » n’a plus rien à voir avec la façon dont fonctionne le cerveau, mais les neurosciences computationnelles, elles continuent, notamment pour aider à faire des prothèses adaptées comme les BCI, Brain Computer Interfaces

Désormais, faire de la science consiste à essayer de comprendre comment fonctionne le monde naturel depuis un modèle. Jusqu’à récemment, on pensait qu’il fallait des théories pour savoir quoi faire des données, mais depuis l’avènement des traitements probabilistes et du Big Data, les modèles théoriques sont devenus inutiles, comme l’expliquait Chris Anderson dans The End of Theory en 2008. En 2017, des chercheurs se sont tout de même demandé si l’on pouvait renverser l’analogie cerveau-ordinateur en tentant de comprendre le fonctionnement d’un microprocesseur depuis les outils des neurosciences. Malgré l’arsenal d’outils à leur disposition, les chercheurs qui s’y sont essayé ont été incapables de produire un modèle de son fonctionnement. Cela nous montre que comprendre un fonctionnement ne nécessite pas seulement des informations techniques ou des données, mais avant tout des concepts pour les organiser. En fait, avoir accès à une quantité illimitée de données ne suffit pas à comprendre ni le processeur ni le cerveau. En 1974, le philosophe des sciences, Thomas Nagel, avait proposé une expérience de pensée avec son article « Quel effet ça fait d’être une chauve-souris ? ». Même si l’on connaissait tout d’une chauve-souris, on ne pourra jamais savoir ce que ça fait d’être une chauve-souris. Cela signifie qu’on ne peut jamais atteindre la vie intérieure d’autrui. Que la subjectivité des autres nous échappe toujours. C’est là le difficile problème de la conscience. 

Albert Moukheiber sur la scène d’USI 2025.

La subjectivité nous échappe

Une émotion désigne trois choses distinctes, rappelle Albert Moukheiber. C’est un état biologique qu’on peut tenter d’objectiver en trouvant des modalités de mesure, comme le tonus musculaire. C’est un concept culturel qui a des ancrages et valeurs très différentes d’une culture l’autre. Mais c’est aussi et d’abord un ressenti subjectif. Ainsi, par exemple, le fait de se sentir triste n’est pas mesurable. « On peut parfaitement comprendre le cortex moteur et visuel, mais on ne comprend pas nécessairement ce qu’éprouve le narrateur de Proust quand il mange la fameuse madeleine. Dix personnes peuvent être émues par un même coucher de soleil, mais sont-elles émues de la même manière ? » 

Notre réductionnisme objectivant est là confronté à des situations qu’il est difficile de mesurer. Ce qui n’est pas sans poser problèmes, notamment dans le monde de l’entreprise comme dans celui de la santé mentale. 

Le monde de l’entreprise a créé d’innombrables indicateurs pour tenter de mesurer la performance des salariés et collaborateurs. Il n’est pas le seul, s’amuse le chercheur sur scène. Les notes des étudiants leurs rappellent que le but est de réussir les examens plus que d’apprendre. C’est la logique de la loi de Goodhart : quand la mesure devient la cible, elle n’est plus une bonne mesure. Pour obtenir des bonus financiers liés au nombre d’opérations réussies, les chirurgiens réalisent bien plus d’opérations faciles que de compliquées. Quand on mesure les humains, ils ont tendance à modifier leur comportement pour se conformer à la mesure, ce qui n’est pas sans effets rebond, à l’image du célèbre effet cobra, où le régime colonial britannique offrit une prime aux habitants de Delhi qui rapporteraient des cobras morts pour les éradiquer, mais qui a poussé à leur démultiplication pour toucher la prime. En entreprises, nombre de mesures réalisées perdent ainsi très vite de leur effectivité. Moukheiber rappelle que les innombrables tests de personnalité ne valent pas mieux qu’un horoscope. L’un des tests le plus utilisé reste le MBTI qui a été développé dans les années 30 par des personnes sans aucune formation en psychologie. Non seulement ces tests n’ont aucun cadre théorique (voir ce que nous en disait le psychologue Alexandre Saint-Jevin, il y a quelques années), mais surtout, « ce sont nos croyances qui sont déphasées. Beaucoup de personnes pensent que la personnalité des individus serait centrale dans le cadre professionnel. C’est oublier que Steve Jobs était surtout un bel enfoiré ! », comme nombre de ces « grands » entrepreneurs que trop de gens portent aux nuesComme nous le rappelions nous-mêmes, la recherche montre en effet que les tests de personnalités peinent à mesurer la performance au travail et que celle-ci a d’ailleurs peu à voir avec la personnalité. « Ces tests nous demandent d’y répondre personnellement, quand ce devrait être d’abord à nos collègues de les passer pour nous », ironise Moukheiber. Ils supposent surtout que la personnalité serait « stable », ce qui n’est certainement pas si vrai. Enfin, ces tests oublient que bien d’autres facteurs ont peut-être bien plus d’importance que la personnalité : les compétences, le fait de bien s’entendre avec les autres, le niveau de rémunération, le cadre de travail… Mais surtout, ils ont tous un effet « barnum » : n’importe qui est capable de se reconnaître dedans. Dans ces tests, les résultats sont toujours positifs, même les gens les plus sadiques seront flattés des résultats. Bref, vous pouvez les passer à la broyeuse. 

Dans le domaine de la santé mentale, la mesure de la subjectivité est très difficile et son absence très handicapante. La santé mentale est souvent vue comme une discipline objectivable, comme le reste de la santé. Le modèle biomédical repose sur l’idée qu’il suffit d’ôter le pathogène pour aller mieux. Il suffirait alors d’enlever les troubles mentaux pour enlever le pathogène. Bien sûr, ce n’est pas le cas. « Imaginez un moment, vous êtes une femme brillante de 45 ans, star montante de son domaine, travaillant dans une entreprise où vous êtes très valorisée. Vous êtes débauché par la concurrence, une entreprise encore plus brillante où vous allez pouvoir briller encore plus. Mais voilà, vous y subissez des remarques sexistes permanentes, tant et si bien que vous vous sentez moins bien, que vous perdez confiance, que vous développez un trouble anxieux. On va alors pousser la personne à se soigner… Mais le pathogène n’est ici pas en elle, il est dans son environnement. N’est-ce pas ici ses collègues qu’il faudrait pousser à se faire soigner ? » 

En médecine, on veut toujours mesurer les choses. Mais certaines restent insondables. Pour mesurer la douleur, il existe une échelle de la douleur.

Exemple d’échelle d’évaluation de la douleur.

« Mais deux personnes confrontés à la même blessure ne vont pas l’exprimer au même endroit sur l’échelle de la douleur. La douleur n’est pas objectivable. On ne peut connaître que les douleurs qu’on a vécu, à laquelle on les compare ». Mais chacun a une échelle de comparaison différente, car personnelle. « Et puis surtout, on est très doué pour ne pas croire et écouter les gens. C’est ainsi que l’endométriose a mis des années pour devenir un problème de santé publique. Une femme à 50% de chance d’être qualifiée en crise de panique quand elle fait un AVC qu’un homme »… Les exemples en ce sens sont innombrables. « Notre obsession à tout mesurer finit par nier l’existence de la subjectivité ». Rapportée à moi, ma douleur est réelle et handicapante. Rapportée aux autres, ma douleur n’est bien souvent perçue que comme une façon de se plaindre. « Les sciences cognitives ont pourtant besoin de meilleures approches pour prendre en compte cette phénoménologie. Nous avons besoin d’imaginer les moyens de mesurer la subjectivité et de la prendre plus au sérieux qu’elle n’est »

La science de la subjectivité n’est pas dénuée de tentatives de mesure, mais elles sont souvent balayées de la main, alors qu’elles sont souvent plus fiables que les mesures dites objectives. « Demander à quelqu’un comment il va est souvent plus parlant que les mesures électrodermales qu’on peut réaliser ». Reste que les mesures physiologiques restent toujours très séduisantes que d’écouter un patient, un peu comme quand vous ajoutez une image d’une IRM à un article pour le rendre plus sérieux qu’il n’est. 

*

Pour conclure la journée, Christian Fauré, directeur scientifique d’Octo Technology revenait sur son thème, l’incalculabilité. « Trop souvent, décider c’est calculer. Nos décisions ne dépendraient plus alors que d’une puissance de calcul, comme nous le racontent les chantres de l’IA qui s’empressent à nous vendre la plus puissante. Nos décisions sont-elles le fruit d’un calcul ? Nos modèles d’affaires dépendent-ils d’un calcul ? Au tout début d’OpenAI, Sam Altman promettait d’utiliser l’IA pour trouver un modèle économique à OpenAI. Pour lui, décider n’est rien d’autre que calculer. Et le calcul semble pouvoir s’appliquer à tout. Certains espaces échappent encore, comme vient de le dire Albert Moukheiber. Tout n’est pas calculable. Le calcul ne va pas tout résoudre. Cela semble difficile à croire quand tout est désormais analysé, soupesé, mesuré« . « Il faut qu’il y ait dans le poème un nombre tel qu’il empêche de compter », disait Paul Claudel. Le poème n’est pas que de la mesure et du calcul, voulait dire Claudel. Il faut qu’il reste de l’incalculable, même chez le comptable, sinon à quoi bon faire ces métiers. « L’incalculable, c’est ce qui donne du sens »

« Nous vivons dans un monde où le calcul est partout… Mais il ne donne pas toutes les réponses. Et notamment, il ne donne pas de sens, comme disait Pascal Chabot. Claude Shannon, dit à ses collègues de ne pas donner de sens et de signification dans les données. Turing qui invente l’ordinateur, explique que c’est une procédure univoque, c’est-à-dire qu’elle est reliée à un langage qui n’a qu’un sens, comme le zéro et le un. Comme si finalement, dans cette abstraction pure, réduite à l’essentiel, il était impossible de percevoir le sens ».

Hubert Guillaud

  • ✇404 Media
  • I Tried Pre-Ordering the Trump Phone. The Page Failed and It Charged My Credit Card the Wrong Amount
    On Monday the Trump Organization announced its own mobile service plan and the “​​T1 Phone,” a customized all-gold mobile phone that its creators say will be made in America. I tried to pre-order the phone and pay the $100 downpayment, hoping to test the phone to see what apps come pre-installed, how secure it really is, and what components it includes when it comes out. The website failed, went to an error page, and then charged my credit card the wrong amount of $64.70. I received a confirm
     

I Tried Pre-Ordering the Trump Phone. The Page Failed and It Charged My Credit Card the Wrong Amount

16 juin 2025 à 12:36
I Tried Pre-Ordering the Trump Phone. The Page Failed and It Charged My Credit Card the Wrong Amount

On Monday the Trump Organization announced its own mobile service plan and the “​​T1 Phone,” a customized all-gold mobile phone that its creators say will be made in America. 

I tried to pre-order the phone and pay the $100 downpayment, hoping to test the phone to see what apps come pre-installed, how secure it really is, and what components it includes when it comes out. The website failed, went to an error page, and then charged my credit card the wrong amount of $64.70. I received a confirmation email saying I’ll receive a confirmation when my order has been shipped, but I haven’t provided a shipping address or paid the full $499 price tag. It is the worst experience I’ve ever faced buying a consumer electronic product and I have no idea whether or how I’ll receive the phone.

“Trump Mobile is going to change the game, we’re building on the movement to put America first, and we will deliver the highest levels of quality and service. Our company is based right here in the United States because we know it’s what our customers want and deserve,” Donald Trump Jr., EVP of the Trump Organization, and obviously one of President Trump’s sons, said in a press release announcing Trump Mobile

  • ✇404 Media
  • RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts
    This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.A family in Utah is suing the Republican National Convention for sending unhinged text messages soliciting donations to Donald Trump’s campaign and continuing to text even after they tried to unsubscribe.“From Trump: ALL HELL JUST BROKE LOOSE! I WAS CONVICTED IN A RIGGED TRIAL!” one example text message in the complaint says. “I need you to read thi
     

RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts

16 juin 2025 à 12:22
RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts

This article was produced in collaboration with Court Watch, an independent outlet that unearths overlooked court records. Subscribe to them here.

A family in Utah is suing the Republican National Convention for sending unhinged text messages soliciting donations to Donald Trump’s campaign and continuing to text even after they tried to unsubscribe.

“From Trump: ALL HELL JUST BROKE LOOSE! I WAS CONVICTED IN A RIGGED TRIAL!” one example text message in the complaint says. “I need you to read this NOW” followed by a link to a donation page.

RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts

The complaint, seeking to become a class-action lawsuit and brought by Utah residents Samantha and Cari Johnson, claims that the RNC, through the affiliated small-donations platform WinRed, violates the Utah Telephone and Facsimile Solicitation Act because the law states “[a] telephone solicitor may not make or cause to be made a telephone solicitation to a person who has informed the telephone solicitor, either in writing or orally, that the person does not wish to receive a telephone call from the telephone solicitor.”

The Johnsons claim that the RNC sent Samantha 17 messages from 16 different phone numbers, nine of the messages after she demanded the messages stop 12 times. Cari received 27 messages from 25 numbers, they claim, and she sent 20 stop requests. The National Republican Senatorial Committee, National Republican Congressional Committee, and Congressional Leadership Fund also sent a slew of texts and similarly didn’t stop after multiple requests, the complaint says. 

On its website, WinRed says it’s an “online fundraising platform supported by a united front of the Trump campaign, RNC, NRSC, and NRCC.” 

RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts
A chart from the complaint showing the numbers of times the RNC and others have texted the plaintiffs.

“Defendants’ conduct is not accidental. They knowingly disregard stop requests and purposefully use different phone numbers to make it impossible to block new messages,” the complaint says.

The complaint also cites posts other people have made on X.com complaining about WinRed’s texts. A quick search for WinRed on X today shows many more people complaining about the same issues. 

RNC Sued Over WinRed's Constant 'ALL HELL JUST BROKE LOOSE!' Fundraising Texts

“I’m seriously considering filing a class action lawsuit against @WINRED. The sheer amount of campaign txts I receive is astounding,” one person wrote on X. “I’ve unsubscribed from probably thousands of campaign texts to no avail. The scam is, if you call Winred, they say it’s campaign initiated. Call campaign, they say it’s Winred initiated. I can’t be the only one!”

Last month, Democrats on the House Judiciary, Oversight and Administration Committees asked the Treasury Department to provide evidence of “suspicious transactions connected to a wide range of Republican and President Donald Trump-aligned fundraising platforms” including WinRed, Politico reported.   

In June 2024, a day after an assassination attempt on Trump during a rally in Pennsylvania, WinRed changed its landing page to all-black with the Trump campaign logo and a black-and-white photograph of Trump raising his fist with blood on his face. “I am Donald J. Trump,” text on the page said. “FEAR NOT! I will always love you for supporting me.”

CNN investigated campaign donation text messaging schemes including WinRed in 2024, and found that the elderly were especially vulnerable to the inflammatory, constant messaging from politicians through text messages begging for donations. And Al Jazeera uncovered FEC records showing people were repeatedly overcharged by WinRed, with one person the outlet spoke to claiming he was charged almost $90,000 across six different credit cards despite thinking he’d only donated small amounts occasionally. “Every single text link goes to WinRed, has the option to ‘repeat your donation’ automatically selected, and uses shady tactics and lies to trick you into clicking on the link,” another donor told Al Jazeera in 2024. “Let’s just say I’m very upset with WinRed. In my view, they are deceitful money-grabbing liars.” 

And in 2020, a class action lawsuit against WinRed made similar claims, but was later dismissed.

  • ✇404 Media
  • Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police
    📄This article was primarily reported using public records requests. We are making it available to all readers as a public service. FOIA reporting can be expensive, please consider subscribing to 404 Media to support this work. Or send us a one time donation via our tip jar here.Local police in Oregon casually offered various surveillance services to federal law enforcement officials from the FBI and ICE, and to other state and local police departments, as part of an informal email and meetup gro
     

Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police

16 juin 2025 à 11:10
📄
This article was primarily reported using public records requests. We are making it available to all readers as a public service. FOIA reporting can be expensive, please consider subscribing to 404 Media to support this work. Or send us a one time donation via our tip jar here.
Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police

Local police in Oregon casually offered various surveillance services to federal law enforcement officials from the FBI and ICE, and to other state and local police departments, as part of an informal email and meetup group of crime analysts, internal emails shared with 404 Media show. 

In the email thread, crime analysts from several local police departments and the FBI introduced themselves to each other and made lists of surveillance tools and tactics they have access to and felt comfortable using, and in some cases offered to perform surveillance for their colleagues in other departments. The thread also includes a member of ICE’s Homeland Security Investigations (HSI) and members of Oregon’s State Police. In the thread, called the “Southern Oregon Analyst Group,” some members talked about making fake social media profiles to surveil people, and others discussed being excited to learn and try new surveillance techniques. The emails show both the wide array of surveillance tools that are available to even small police departments in the United States and also shows informal collaboration between local police departments and federal agencies, when ordinarily agencies like ICE are expected to follow their own legal processes for carrying out the surveillance. 

In one case, a police analyst for the city of Medford, Oregon, performed Flock automated license plate reader (ALPR) lookups for a member of ICE’s HSI; later, that same police analyst asked the HSI agent to search for specific license plates in DHS’s own border crossing license plate database. The emails show the extremely casual and informal nature of what partnerships between police departments and federal law enforcement can look like, which may help explain the mechanics of how local police around the country are performing Flock automated license plate reader lookups for ICE and HSI even though neither group has a contract to use the technology, which 404 Media reported last month

Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police
An email showing HSI asking for a license plate lookup from police in Medford, Oregon

Kelly Simon, the legal director for the American Civil Liberties Union of Oregon, told 404 Media “I think it’s a really concerning thread to see, in such a black-and-white way. I have certainly never seen such informal, free-flowing of information that seems to be suggested in these emails.”

In that case, in 2021, a crime analyst with HSI emailed an analyst at the Medford Police Department with the subject line “LPR Check.” The email from the HSI analyst, who is also based in Medford, said they were told to “contact you and request a LPR check on (2) vehicles,” and then listed the license plates of two vehicles. “Here you go,” the Medford Police Department analyst responded with details of the license plate reader lookup. “I only went back to 1/1/19, let me know if you want me to check further back.” In 2024, the Medford police analyst emailed the same HSI agent and told him that she was assisting another police department with a suspected sex crime and asked him to “run plates through the border crossing system,” meaning the federal ALPR system at the Canada-US border. “Yes, I can do that. Let me know what you need and I’ll take a look,” the HSI agent said. 

More broadly, the emails, obtained using a public records request by Information for Public Use, an anonymous group of researchers in Oregon who have repeatedly uncovered documents about government surveillance, reveal the existence of the “Southern Oregon Analyst Group.” The emails span between 2021 and 2024 and show local police eagerly offering various surveillance services to each other as part of their own professional development. 

In a 2023 email thread where different police analysts introduced themselves, they explained to each other what types of surveillance software they had access to, which ones they use the most often, and at times expressed an eagerness to try new techniques. 

Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police

“This is my first role in Law Enforcement, and I've been with the Josephine County Sheriff's Office for 6 months, so I'm new to the game,” an email from a former Pinkerton security contractor to officials at 10 different police departments, the FBI, and ICE, reads. “Some tools I use are Flock, TLO, Leads online, WSIN, Carfax for police, VIN Decoding, LEDS, and sock puppet social media accounts. In my role I build pre-raid intelligence packages, find information on suspects and vehicles, and build link charts showing connections within crime syndicates. My role with [Josephine Marijuana Enforcement Team] is very intelligence and research heavy, but I will do the occasional product with stats. I would love to be able to meet everyone at a Southern Oregon analyst meet-up in the near future. If there is anything I can ever provide anyone from Josephine County, please do not hesitate to reach out!” The surveillance tools listed here include automatic license plate reading technology, social media monitoring tools, people search databases, and car ownership history tools. 

An investigations specialist with the Ashland Police Department messaged the group, said she was relatively new to performing online investigations, and said she was seeking additional experience. “I love being in a support role but worry patrol doesn't have confidence in me. I feel confident with searching through our local cad portal, RMS, Evidence.com, LeadsOnline, carfax and TLO. Even though we don't have cameras in our city, I love any opportunity to search for something through Flock,” she said. “I have much to learn with sneaking around in social media, and collecting accurate reports from what is inputted by our department.”

Emails Reveal the Casual Surveillance Alliance Between ICE and Local Police

A crime analyst with the Medford Police Department introduced themselves to the group by saying “The Medford Police Department utilizes the license plate reader systems, Vigilant and Flock. In the next couple months, we will be starting our transition to the Axon Fleet 3 cameras. These cameras will have LPR as well. If you need any LPR searches done, please reach out to me or one of the other analysts here at MPD. Some other tools/programs that we have here at MPD are: ESRI, Penlink PLX, CellHawk, TLO, LeadsOnline, CyberCheck, Vector Scheduling/CrewSense & Guardian Tracking, Milestone XProtect city cameras, AXON fleet and body cams, Lexipol, HeadSpace, and our RMS is Central Square (in case your agency is looking into purchasing any of these or want more information on them).”

A fourth analyst said “my agency uses Tulip, GeoShield, Flock LPR, LeadsOnline, TLO, Axon fleet and body cams, Lexipol, LEEP, ODMap, DMV2U, RISS/WSIN, Crystal Reports, SSRS Report Builder, Central Square Enterprise RMS, Laserfiche for fillable forms and archiving, and occasionally Hawk Toolbox.” Several of these tools are enterprise software solutions for police departments, which include things like police report management software, report creation software, and stress management and wellbeing software, but many of them are surveillance tools.  

At one point in the 2023 thread, an FBI intelligence analyst for the FBI’s Portland office chimes in, introduces himself, and said “I think I've been in contact with most folks on this email at some point in the past […] I look forward to further collaboration with you all.”

The email thread also planned in-person meetups and a “mini-conference” last year that featured a demo from a company called CrimeiX, a police information sharing tool.  

A member of Information for Public Use told 404 Media “it’s concerning to me to see them building a network of mass surveillance.”

“Automated license plate recognition software technology is something that in and of itself, communities are really concerned about,” the member of Information for Public Use said. “So I think when we combine this very obvious mass surveillance technology with a network of interagency crime analysts that includes local police who are using sock puppet accounts to spy on anyone and their mother and then that information is being pretty freely shared with federal agents, you know, including Homeland Security Investigations, and we see the FBI in the emails as well. It's pretty disturbing.” They added, as we have reported before, that many of these technologies were deployed under previous administrations but have become even more alarming when combined with the fact that the Trump administration has changed the priorities of ICE and Homeland Security Investigations. 

“The whims of the federal administration change, and this technology can be pointed in any direction,” they said. “Local law enforcement might be justifying this under the auspices of we're fighting some form of organized crime, but one of the crimes HSI investigates is work site enforcement investigations, which sound exactly like the kind of raids on workplaces that like the country is so upset about right now.”

Simon, of ACLU Oregon, said that such informal collaboration is not supposed to be happening in Oregon.

“We have, in Oregon, a lot of really strong protections that ensure that our state resources, including at the local level, are not going to support things that Oregonians disagree with or have different values around,” she said. “Oregon has really strong firewalls between local resources, and federal resources or other state resources when it comes to things like reproductive justice or immigrant justice. We have really strong shield laws, we have really strong sanctuary laws, and when I see exchanges like this, I’m very concerned that our firewalls are more like sieves because of this kind of behind-the-scenes, lax approach to protecting the data and privacy of Oregonians.”

Simon said that collaboration between federal and local cops on surveillance should happen “with the oversight of the court. Getting a warrant to request data from a local agency seems appropriate to me, and it ensures there’s probable cause, that the person whose information is being sought is sufficiently suspected of a crime, and that there are limits to the scope, about of information that's being sought and specifics about what information is being sought. That's the whole purpose of a warrant.”

Over the last several weeks, our reporting has led multiple municipalities to reconsider how the license plate reading technology Flock is used, and it has spurred an investigation by the Illinois Secretary of State office into the legality of using Flock cameras in the state for immigration-related searches, because Illinois specifically forbids local police from assisting federal police on immigration matters.

404 Media contacted all of the police departments on the Southern Oregon Analyst Group for comment and to ask them about any guardrails they have for the sharing of surveillance tools across departments or with the federal government. Geoffrey Kirkpatrick, a lieutenant with the Medford Police Department, said the group is “for professional networking and sharing professional expertise with each other as they serve their respective agencies.” 

“The Medford Police Department’s stance on resource-sharing with ICE is consistent with both state law and federal law,” Kirkpatrick said. “The emails retrieved for that 2025 public records request showed one single instance of running LPR information for a Department of Homeland Security analyst in November 2021. Retrieving those files from that single 2021 matter to determine whether it was an DHS case unrelated to immigration, whether a criminal warrant existed, etc would take more time than your publication deadline would allow, and the specifics of that one case may not be appropriate for public disclosure regardless.” (404 Media reached out to Medford Police Department a week before this article was published). 

A spokesperson for the Central Point Police Department said it “utilizes technology as part of investigations, we follow all federal, state, and local law regarding use of such technology and sharing of any such information. Typically we do not use our tools on behalf of other agencies.”

A spokesperson for Oregon’s Department of Justice said it did not have comment and does not participate in the group. The other police departments in the group did not respond to our request for comment.

  • ✇404 Media
  • Meta Users Feel Less Safe Since It Weakened ‘Hateful Conduct’ Policy, Survey Finds
    A survey of 7,000 Facebook, Instagram, and Threads users found that most people feel less safe on Meta’s platforms since CEO Mark Zuckerberg abandoned fact-checking in January.The report, written by Jenna Sherman at UltraViolet, Ana Clara-Toledo at All Out, and Leanna Garfield at GLAAD, surveyed people who belong to what Meta refers to as “protected characteristic groups,” which include “people targeted based on their race, ethnicity, national origin, disability, religious affiliation, caste,
     

Meta Users Feel Less Safe Since It Weakened ‘Hateful Conduct’ Policy, Survey Finds

16 juin 2025 à 07:00
Meta Users Feel Less Safe Since It Weakened ‘Hateful Conduct’ Policy, Survey Finds

A survey of 7,000 Facebook, Instagram, and Threads users found that most people feel less safe on Meta’s platforms since CEO Mark Zuckerberg abandoned fact-checking in January.

The report, written by Jenna Sherman at UltraViolet, Ana Clara-Toledo at All Out, and Leanna Garfield at GLAAD, surveyed people who belong to what Meta refers to as “protected characteristic groups,” which include “people targeted based on their race, ethnicity, national origin, disability, religious affiliation, caste, sexual orientation, sex, gender identity, or serious disease,” the report says. The average age of respondents was 50 years, and the survey asked them to respond to questions including “How well do you feel Meta’s new policy changes protect you and all users from being exposed to or targeted by harmful content?” and “Have you been the target of any form of harmful content on any Meta platform since January 2025?” 

One in six of respondents reported being targeted with gender-based or sexual violence on Meta platforms, and 66 percent of respondents said they’ve witnessed harmful content on Meta platforms. The survey defined harmful content as “content that involves direct attacks against people based on a protected characteristic.”  

Almost all of the users surveyed—more than 90 percent—said they’re concerned about increasing harmful content, and feel less protected from being exposed to or targeted by harmful content on Meta’s platforms.

“I have seen an extremely large influx of hate speech directed towards many different marginalized groups since Jan. 2025,” one user wrote in the comments section of the survey. “I have also noted a large increase in ‘fake pages’ generating false stories to invoke an emotional response from people who are clearly against many marginalized groups since Jan. 2025.”

“I rarely see friends’ posts [now], I am exposed to obscene faked sexual images in the opening boxes, I am battered with commercial ads for products that are crap,” another wrote, adding that they were moving to Bluesky and Substack for “less gross posts.”

404 Media has extensively reported on the kinds of gruesome slop these users are referring to. Meta’s platforms allow AI-generated spam schemes to run rampant, at the expense of human-made, quality content. 

In January, employees at Meta told 404 Media in interviews and demonstrated with leaked internal conversations that people working there were furious about the changes. A member of the public policy team said in Meta’s internal workspace that the changes to the Hateful Conduct policy—to allow users to call gay people “mentally ill” and immigrants “trash,” for example—was simply an effort to “undo mission creep.” “Reaffirming our core value of free expression means that we might see content on our platforms that people find offensive … yesterday’s changes not only open up conversation about these subjects, but allow for counterspeech on what matters to users,” the policy person said in a thread addressing angry Meta employees.

Zuckerberg has increasingly chosen to pander to the Trump administration through public support and moderation slackening on his platforms. In the January announcement, he promised to “get rid of a bunch of restrictions on topics like immigration and gender that are just out of touch with mainstream discourse.” In practice, according to leaked internal documents, that meant allowing violent hate speech on his platforms, including sexism, racism, and bigotry.

Several respondents to the survey wrote that the changes have resulted in a hostile social media environment. “I was told that as a woman I should be ‘properly fucked by a real man’ to ‘fix my head’ regarding gender equality and LGBT+ rights,” one said.“I’ve been told women should know their place if we want to support America. I’ve been sent DMs requesting contact based on my appearance. I’ve been primarily stalked due to my political orientation,” another wrote. Studies show that rampant hate speech online can predict real-world violence.

The authors of the report wrote that they want to see Meta hire an independent third-party to “formally analyze changes in harmful content facilitated by the policy changes” made in January, and for the social media giant to bring back the moderation standards that were in place before then. But all signs point to Zuckerberg not just liking the content on his site that makes it worse, but ignoring the issue completely to build more harmful chatbots and spend billions of dollars on a “superintelligence” project.

  • ✇Dans les algorithmes
  • “Il est probable que l’empreinte environnementale de l’IA soit aujourd’hui la plus faible jamais atteinte”
    Alors que l’IA s’intègre peu à peu partout dans nos vies, les ressources énergétiques nécessaires à cette révolution sont colossales. Les plus grandes entreprises technologiques mondiales l’ont bien compris et ont fait de l’exploitation de l’énergie leur nouvelle priorité, à l’image de Meta et Microsoft qui travaillent à la mise en service de centrales nucléaires pour assouvir leurs besoins. Tous les Gafams ont des programmes de construction de data centers démesurés avec des centaines de millia
     

“Il est probable que l’empreinte environnementale de l’IA soit aujourd’hui la plus faible jamais atteinte”

16 juin 2025 à 01:00

Alors que l’IA s’intègre peu à peu partout dans nos vies, les ressources énergétiques nécessaires à cette révolution sont colossales. Les plus grandes entreprises technologiques mondiales l’ont bien compris et ont fait de l’exploitation de l’énergie leur nouvelle priorité, à l’image de Meta et Microsoft qui travaillent à la mise en service de centrales nucléaires pour assouvir leurs besoins. Tous les Gafams ont des programmes de construction de data centers démesurés avec des centaines de milliards d’investissements, explique la Technology Review. C’est le cas par exemple à Abilene au Texas, où OpenAI (associé à Oracle et SoftBank) construit un data center géant, premier des 10 mégasites du projet Stargate, explique un copieux reportage de Bloomberg, qui devrait coûter quelque 12 milliards de dollars (voir également le reportage de 40 minutes en vidéo qui revient notamment sur les tensions liées à ces constructions). Mais plus que de centres de données, il faut désormais parler « d’usine à IA », comme le propose le patron de Nvidia, Jensen Huang. 

“De 2005 à 2017, la quantité d’électricité destinée aux centres de données est restée relativement stable grâce à des gains d’efficacité, malgré la construction d’une multitude de nouveaux centres de données pour répondre à l’essor des services en ligne basés sur le cloud, de Facebook à Netflix”, explique la TechReview. Mais depuis 2017 et l’arrivée de l’IA, cette consommation s’est envolée. Les derniers rapports montrent que 4,4 % de l’énergie totale aux États-Unis est désormais destinée aux centres de données. “Compte tenu de l’orientation de l’IA – plus personnalisée, capable de raisonner et de résoudre des problèmes complexes à notre place, partout où nous regardons –, il est probable que notre empreinte IA soit aujourd’hui la plus faible jamais atteinte”. D’ici 2028, l’IA à elle seule pourrait consommer chaque année autant d’électricité que 22 % des foyers américains.

“Les chiffres sur la consommation énergétique de l’IA court-circuitent souvent le débat, soit en réprimandant les comportements individuels, soit en suscitant des comparaisons avec des acteurs plus importants du changement climatique. Ces deux réactions esquivent l’essentiel : l’IA est incontournable, et même si une seule requête est à faible impact, les gouvernements et les entreprises façonnent désormais un avenir énergétique bien plus vaste autour des besoins de l’IA”. ChatGPT est désormais considéré comme le cinquième site web le plus visité au monde, juste après Instagram et devant X. Et ChatGPT n’est que l’arbre de la forêt des applications de l’IA qui s’intègrent partout autour de nous. Or, rappelle la Technology Review, l’information et les données sur la consommation énergétique du secteur restent très parcellaires et lacunaires. Le long dossier de la Technology Review rappelle que si l’entraînement des modèles est énergétiquement coûteux, c’est désormais son utilisation qui devient problématique, notamment, comme l’explique très pédagogiquement Le Monde, parce que les requêtes dans un LLM, recalculent en permanence ce qu’on leur demande (et les calculateurs qui évaluent la consommation énergétique de requêtes selon les moteurs d’IA utilisés, comme Ecologits ou ComparIA s’appuient sur des estimations). Dans les 3000 centres de données qu’on estime en activité aux Etats-Unis, de plus en plus d’espaces sont consacrés à des infrastructures dédiées à l’IA, notamment avec des serveurs dotés de puces spécifiques qui ont une consommation énergétique importante pour exécuter leurs opérations avancées sans surchauffe.

Calculer l’impact énergétique d’une requête n’est pas aussi simple que de mesurer la consommation de carburant d’une voiture, rappelle le magazine. “Le type et la taille du modèle, le type de résultat généré et d’innombrables variables indépendantes de votre volonté, comme le réseau électrique connecté au centre de données auquel votre requête est envoyée et l’heure de son traitement, peuvent rendre une requête mille fois plus énergivore et émettrice d’émissions qu’une autre”. Outre cette grande variabilité de l’impact, il faut ajouter l’opacité des géants de l’IA à communiquer des informations et des données fiables et prendre en compte le fait que nos utilisations actuelles de l’IA sont bien plus frustres que les utilisations que nous aurons demain, dans un monde toujours plus agentif et autonome. La taille des modèles, la complexité des questions sont autant d’éléments qui influent sur la consommation énergétique. Bien évidemment, la production de vidéo consomme plus d’énergie qu’une production textuelle. Les entreprises d’IA estiment cependant que la vidéo générative a une empreinte plus faible que les tournages et la production classique, mais cette affirmation n’est pas démontrée et ne prend pas en compte l’effet rebond que génèrerait les vidéos génératives si elles devenaient peu coûteuses à produire. 

La Techno Review propose donc une estimation d’usage quotidien, à savoir en prenant comme moyenne le fait de poser 15 questions à un modèle d’IA génératives, faire 10 essais d’image et produire 5 secondes de vidéo. Ce qui équivaudrait (très grossièrement) à consommer 2,9 kilowattheures d’électricité, l’équivalent d’un micro-onde allumé pendant 3h30. Ensuite, les journalistes tentent d’évaluer l’impact carbone de cette consommation qui dépend beaucoup de sa localisation, selon que les réseaux sont plus ou moins décarbonés, ce qui est encore bien peu le cas aux Etats-Unis (voir notamment l’explication sur les modalités de calcul mobilisées par la Tech Review). “En Californie, produire ces 2,9 kilowattheures d’électricité produirait en moyenne environ 650 grammes de dioxyde de carbone. Mais produire cette même électricité en Virginie-Occidentale pourrait faire grimper le total à plus de 1 150 grammes”. On peut généraliser ces estimations pour tenter de calculer l’impact global de l’IA… et faire des calculs compliqués pour tenter d’approcher la réalité… “Mais toutes ces estimations ne reflètent pas l’avenir proche de l’utilisation de l’IA”. Par exemple, ces estimations reposent sur l’utilisation de puces qui ne sont pas celles qui seront utilisées l’année prochaine ou la suivante dans les “usines à IA” que déploie Nvidia, comme l’expliquait son patron, Jensen Huang, dans une des spectaculaires messes qu’il dissémine autour du monde. Dans cette course au nombre de token générés par seconde, qui devient l’indicateur clé de l’industrie, c’est l’architecture de l’informatique elle-même qui est modifiée. Huang parle de passage à l’échelle qui nécessite de générer le plus grand nombre de token possible et le plus rapidement possible pour favoriser le déploiement d’une IA toujours plus puissante. Cela passe bien évidemment par la production de puces et de serveurs toujours plus puissants et toujours plus efficaces. 

« Dans ce futur, nous ne nous contenterons pas de poser une ou deux questions aux modèles d’IA au cours de la journée, ni de leur demander de générer une photo”. L’avenir, rappelle la Technology Review, est celui des agents IA effectuent des tâches pour nous, où nous discutons en continue avec des agents, où nous “confierons des tâches complexes à des modèles de raisonnement dont on a constaté qu’ils consomment 43 fois plus d’énergie pour les problèmes simples, ou à des modèles de « recherche approfondie”, qui passeront des heures à créer des rapports pour nous ». Nous disposerons de modèles d’IA “personnalisés” par l’apprentissage de nos données et de nos préférences. Et ces modèles sont appelés à s’intégrer partout, des lignes téléphoniques des services clients aux cabinets médicaux… Comme le montrait les dernières démonstrations de Google en la matière : “En mettant l’IA partout, Google souhaite nous la rendre invisible”. “Il ne s’agit plus de savoir qui possède les modèles les plus puissants, mais de savoir qui les transforme en produits performants”. Et de ce côté, là course démarre à peine. Google prévoit par exemple d’intégrer l’IA partout, pour créer des résumés d’email comme des mailings automatisés adaptés à votre style qui répondront pour vous. Meta imagine intégrer l’IA dans toute sa chaîne publicitaire pour permettre à quiconque de générer des publicités et demain, les générer selon les profils : plus personne ne verra la même ! Les usages actuels de l’IA n’ont rien à voir avec les usages que nous aurons demain. Les 15 questions, les 10 images et les 5 secondes de vidéo que la Technology Review prend comme exemple d’utilisation quotidienne appartiennent déjà au passé. Le succès et l’intégration des outils d’IA des plus grands acteurs que sont OpenAI, Google et Meta vient de faire passer le nombre estimé des utilisateurs de l’IA de 700 millions en mars à 3,5 milliards en mai 2025

”Tous les chercheurs interrogés ont affirmé qu’il était impossible d’appréhender les besoins énergétiques futurs en extrapolant simplement l’énergie utilisée par les requêtes d’IA actuelles.” Le fait que les grandes entreprises de l’IA se mettent à construire des centrales nucléaires est d’ailleurs le révélateur qu’elles prévoient, elles, une explosion de leurs besoins énergétiques. « Les quelques chiffres dont nous disposons peuvent apporter un éclairage infime sur notre situation actuelle, mais les années à venir sont incertaines », déclare Sasha Luccioni de Hugging Face. « Les outils d’IA générative nous sont imposés de force, et il devient de plus en plus difficile de s’en désengager ou de faire des choix éclairés en matière d’énergie et de climat. »

La prolifération de l’IA fait peser des perspectives très lourdes sur l’avenir de notre consommation énergétique. “Entre 2024 et 2028, la part de l’électricité américaine destinée aux centres de données pourrait tripler, passant de 4,4 % actuellement à 12 %” Toutes les entreprises estiment que l’IA va nous aider à découvrir des solutions, que son efficacité énergétique va s’améliorer… Et c’est effectivement le cas. A entendre Jensen Huang de Nvidia, c’est déjà le cas, assure-t-il en vantant les mérites des prochaines génération de puces à venir. Mais sans données, aucune “projection raisonnable” n’est possible, estime les contributeurs du rapport du département de l’énergie américain. Surtout, il est probable que ce soient les usagers qui finissent par en payer le prix. Selon une nouvelle étude, les particuliers pourraient finir par payer une partie de la facture de cette révolution de l’IA. Les chercheurs de l’Electricity Law Initiative de Harvard ont analysé les accords entre les entreprises de services publics et les géants de la technologie comme Meta, qui régissent le prix de l’électricité dans les nouveaux centres de données gigantesques. Ils ont constaté que les remises accordées par les entreprises de services publics aux géants de la technologie peuvent augmenter les tarifs d’électricité payés par les consommateurs. Les impacts écologiques de l’IA s’apprêtent donc à être maximums, à mesure que ses déploiements s’intègrent partout. “Il est clair que l’IA est une force qui transforme non seulement la technologie, mais aussi le réseau électrique et le monde qui nous entoure”.

L’article phare de la TechReview, se prolonge d’un riche dossier. Dans un article, qui tente de contrebalancer les constats mortifères que le magazine dresse, la TechReview rappelle bien sûr que les modèles d’IA vont devenir plus efficaces, moins chers et moins gourmands énergétiquement, par exemple en entraînant des modèles avec des données plus organisées et adaptées à des tâches spécifiques. Des perspectives s’échaffaudent aussi du côté des puces et des capacités de calculs, ou encore par l’amélioration du refroidissement des centres de calculs. Beaucoup d’ingénieurs restent confiants. “Depuis, l’essor d’internet et des ordinateurs personnels il y a 25 ans, à mesure que la technologie à l’origine de ces révolutions s’est améliorée, les coûts de l’énergie sont restés plus ou moins stables, malgré l’explosion du nombre d’utilisateurs”. Pas sûr que réitérer ces vieilles promesses suffise. 

Comme le disait Gauthier Roussilhe, nos projections sur les impacts environnementaux à venir sont avant toutes coincées dans le présent. Et elles le sont d’autant plus que les mesures de la consommation énergétique de l’IA sont coincées dans les mesures d’hier, sans être capables de prendre en compte l’efficience à venir et que les effets rebonds de la consommation, dans la perspective de systèmes d’IA distribués partout, accessibles partout, voire pire d’une IA qui se substitue à tous les usages numériques actuels, ne permettent pas d’imaginer ce que notre consommation d’énergie va devenir. Si l’efficience énergétique va s’améliorer, le rebond des usages par l’intégration de l’IA partout, lui, nous montre que les gains obtenus sont toujours totalement absorbés voir totalement dépassés avec l’extension et l’accroissement des usages. 

  • ✇404 Media
  • Climate Change Warps Brains in the Womb, Scientists Discover
    Welcome back to the Abstract! This week, it’s time for a walk in the woods. These particular woods have been dead and buried for centuries, mind you, but they still have a lot to say about the tumultuous events they experienced across thousands of years.Then: exposure to climate change starts in the womb; CYBORG TADPOLES; get swole with this new dinosaur diet; the long march of an ancestral reptile; and, finally, pregaming for science. The saga of the sunken cypress Napora, Katharine et al. “
     

Climate Change Warps Brains in the Womb, Scientists Discover

14 juin 2025 à 09:00
Climate Change Warps Brains in the Womb, Scientists Discover

Welcome back to the Abstract! 

This week, it’s time for a walk in the woods. These particular woods have been dead and buried for centuries, mind you, but they still have a lot to say about the tumultuous events they experienced across thousands of years.

Then: exposure to climate change starts in the womb; CYBORG TADPOLES; get swole with this new dinosaur diet; the long march of an ancestral reptile; and, finally, pregaming for science. 

The saga of the sunken cypress 

Napora, Katharine et al. “Subfossil bald cypress trees suggest localized, enduring effects of major climatic episodes on the Southeast Atlantic Coast of the United States.” Proceedings of the National Academy of Sciences.

For thousands of years, a forest filled with bald cypress trees thrived in coastal Georgia. But climate shifts caused by volcanic eruptions and a possible comet impact wreaked havoc on this environment, eventually leading to the death of these ancient woods by the year 1600.

Now scientists have exhumed dozens of the magnificent trees, which were buried at the mouth of the Altamaha River for centuries. The dead trees are well-preserved as subfossils, meaning they are only partially fossilized, allowing researchers to count tree rings, conduct radiocarbon dating, and reconstruct the epic tale of this long-lived grove.

“This is the largest intact deposit of subfossil Holocene cypress trees ever analyzed in the literature from the Southeast United States…with specimens spanning almost six millennia,” said researchers led by Katharine Napora of Florida Atlantic University in their study. 

In ideal conditions, bald cypress trees can live for millennia; for instance, one tree known as the Senator in Longwood, Florida was about 3,500-years-old when it died in a 2012 fire. But Napora’s team found that their subfossil trees experienced a collapse in life expectancy during the Vandal Minimum (VM) environmental downturn, which began around 500 CE. Trees that sprouted after this event only lived about half as long as those born before, typically under 200 years. 

 

Climate Change Warps Brains in the Womb, Scientists Discover
Study authors Katharine Napora and Craig Jacobs with an ancient cypress tree near the Georgia coast. Image: Florida Atlantic University

The reasons for this downturn are potentially numerous, including volcanic eruptions and a possible comet strike. The researchers say that tree-ring evidence shows “a reduction in solar radiation in 536 and 541 to 544 CE, likely the consequence of a volcanic dust veil…Greenlandic ice cores also contain particles rich in elements suggesting dust originating from a comet, dating to 533 to 540 CE.”

The possibility that a comet struck Earth at this time has been debated for decades, but many scientists think that volcanic eruptions can account for the extreme cooling without invoking space rocks. In any case, the world was rocked by a series of unfortunate events that produced a variety of localized impacts.This Georgian tree cemetery presents a new record of those tumultuous times which “speaks to the long-term impacts of major climatic episodes in antiquity” and “underscores the vulnerability of 21st-century coastal ecosystems to the destabilizing effects of large-scale climatic downturns,” according to the study.

In other news…

PSA: climate risks begin before you’re born 

DeIngeniis, Donato et al. “Prenatal exposure to extreme ambient heat may amplify the adverse impact of Superstorm Sandy on basal ganglia volume among school-aged children.” PLOS One.

In addition to disrupting long-lived trees, climate change poses a threat to people—starting in the womb. A new study tracked the brain development of children whose mothers endured Superstorm Sandy while pregnant, revealing that prenatal exposure to extreme weather events affect neural and emotional health.

“Prenatal exposure to Superstorm Sandy impacted child brain development,” said researchers led by Donato DeIngeniis of the City University of New York. The team found that a group of 8-year-old children whose mothers experienced the 2012 disaster while pregnant had noticeable differences in their basal ganglia, a brain region involved in motor skills and emotional regulation. 

Exposure to both the hurricane and associated extreme heat (defined as temperatures above 95°F) was linked to both a larger pallidum and smaller nucleus accumbens, both subregions of the basal ganglia, compared to unexposed peers. The findings hint at a higher risk of emotional and behavioral disruption, or other impairments, as a consequence of exposure in the womb, but the study said more research is necessary to confirm those associations. 

“Extreme weather events and natural disasters are projected to increase in frequency and magnitude. In addition to promoting initiatives to combat climate change, it is imperative to alert pregnant individuals to the ongoing danger of exposure to extreme climate events,” the team said.

Here come the cyborg tadpoles

Sheng, Hao, Liu, Ren, Li, Qiang et al. “Brain implantation of soft bioelectronics via embryonic development.” Nature.

Scientists have a long tradition of slapping sensors onto brains to monitor whatever the heck is going on in there. The latest edition: Cyborg tadpoles. 

By implanting a microelectrode array into embryonic frogs and axolotls, a team of researchers was able to track neural development and record brain activity with no detectable adverse effects on the tadpoles.

Climate Change Warps Brains in the Womb, Scientists Discover
The cyborg tadpoles in question. Image: Liu Lab / Harvard SEAS

“Cyborg tadpoles showed normal development through later stages, showing comparable morphology, survival rates and developmental timing to control tadpoles,” said researchers co-led by Hao Sheng, Ren Liu, and Qiang Li of Harvard University. “Future combination of this system with virtual-reality platforms could provide a powerful tool for investigating behaviour- and sensory-specific brain activity during development.” 

The future didn’t deliver personal jetpacks, but we may get virtual-reality tours of amphibian cyborg brains, so there’s that. 

We finally know for certain what sauropods ate 

Poropat, Stephen F. et al. “Fossilized gut contents elucidate the feeding habits of sauropod dinosaurs.” Current Biology.

Once upon a time, a long-necked sauropod dinosaur from the Diamantinasaurus family was chowing down on a variety of plants. Shortly afterward, it died (RIP). 100 million years later, this leafy last meal has now provided the first direct evidence that sauropods—the largest animals ever to walk on land—were herbivores. 

Climate Change Warps Brains in the Womb, Scientists Discover
Fossilized ferns, conifers, and other plants were found in the Australian Diamantinasaurus cololite. Image: Stephen Poropat

“Gut contents for sauropod dinosaurs—perhaps the most ecologically impactful terrestrial herbivores worldwide throughout much of the Jurassic and Cretaceous, given their gigantic sizes—have remained elusive,” said researchers led by Stephen Poropat of Curtin University. “The Diamantinasaurus cololite (fossilized gut contents) described herein provides the first direct, empirical support for the long-standing hypothesis of sauropod herbivory.”

Scientists have long assumed that sauropods were veggie-saurs based on their anatomy, but it’s cool to finally have confirmation by looking in the belly of this beast. 

Life finds a way through the “dead zone”

Flannery-Sutherland, Joseph et al. “Landscape-explicit phylogeography illuminates the ecographic radiation of early archosauromorph reptiles.” Nature Ecology & Evolution.

Birds, crocodiles, and dinosaurs are all descended from an ancestral lineage of reptiles called archosauromorphs. These troopers managed to survive Earth’s most devastating extinction event, called the end-Permian or “Great Dying,” a global warming catastrophe that wiped out more than half of all land animals and 81 percent of marine life some 250 million years ago.

Climate Change Warps Brains in the Womb, Scientists Discover
Children of the archosaurs, chillin’. Image: Timothy A. Gonsalves

Now, paleontologists have found clues indicating how they succeeded by reconstructing archosauromorph dispersal patterns with models of ancient landscapes and evolutionary trees. The results suggest that these animals endured 10,000-mile marches through “tropical dead zones.”

These archosauromorph “dispersals through the Pangaean tropical dead zone…contradict its perception as a hard barrier to vertebrate movement,” said researchers led by Joseph Flannery-Sutherland of the University of Birmingham. “This remarkable tolerance of climatic adversity was probably integral to their later evolutionary success.”

The science of spectator sports

Xygalatas, Dimitris et al “Route of fire: Pregame rituals and emotional synchrony among Brazilian football fans.” Proceedings of the National Academy of Sciences.

In Brazil, football fans participate in a pregame ritual known as the Rua de Fogo, or Street of Fire. As buses carrying teams arrive at the stadium, fans greet the players with flares, smoke bombs, fireworks, flags, cheers, and chants.

Now, scientists have offered a glimpse into the ecstatic emotions of these crowds by enlisting  17 fans, including a team bus driver, to wear heart rate monitors in advance of a state championship final  between local teams. The results showed that fans’ heart rates synced up during periods of “emotional synchrony.” 

“We found that the Rua de Fogo ritual preceding the football match exhibited particularly high levels of emotional synchrony—surpassing even those observed during the game itself, which was among the season’s most important,” said researchers led by Dimitris Xygalatas of the University of Connecticut. “These findings suggest that fan rituals play important roles in fostering shared emotional experiences, reinforcing the broader appeal of sports as a site of social connection and identity formation.”

Wishing everyone an emotionally synchronous weekend! Thanks for reading and see you next week.

  • ✇404 Media
  • Behind the Blog: Advertising and Aircraft
    This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss advertising, protests, and aircraft.EMANUEL: On Thursday Meta announced that it has filed a lawsuit in Hong Kong against Joy Timeline HK Limited, the company that operates a popular nudify app called Crush that we have covered previously. Meta’s position is that it hasn’t been able to prevent Crush from advertising its nudify app on its pla
     

Behind the Blog: Advertising and Aircraft

13 juin 2025 à 11:58
Behind the Blog: Advertising and Aircraft

This is Behind the Blog, where we share our behind-the-scenes thoughts about how a few of our top stories of the week came together. This week, we discuss advertising, protests, and aircraft.

EMANUEL: On Thursday Meta announced that it has filed a lawsuit in Hong Kong against Joy Timeline HK Limited, the company that operates a popular nudify app called Crush that we have covered previously

Meta’s position is that it hasn’t been able to prevent Crush from advertising its nudify app on its platform despite it violating its policies because Crush is “highly adversarial” and “constantly evolving their tactics to avoid enforcement.” We’ve seen Crush and other nudify apps create hundreds of Meta advertising accounts and different domain names that all link back to the same service in order to avoid detection. If Meta bans an advertising account or URL, Crush simply creates another. In theory, Meta always has ways of detecting if an ad contains nudity, but nudify apps can easily circumvent those measures as well. As I say in my post about the lawsuit, Meta still hasn’t explained why it appears to have different standards for content in ads versus regular posts on its platform, but there’s no doubt that it does take action against nudify ads when it’s easy for it do so, and that these nudify ads are actively trying to avoid Meta’s moderation when it does attempt to get rid of them. 

  • ✇404 Media
  • Humans Have Now Seen the Dawn of Time from Earth After Breakthrough
    🌘Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week. Scientists have captured an unprecedented glimpse of cosmic dawn, an era more than 13 billion years ago, using telescopes on the surface of the Earth. This marks the first time humans have seen signatures of the first stars interacting with the early universe from our planet, rather than space. This ancient epoch when the first stars lit up the universe has
     

Humans Have Now Seen the Dawn of Time from Earth After Breakthrough

12 juin 2025 à 10:14
🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
Humans Have Now Seen the Dawn of Time from Earth After Breakthrough

Scientists have captured an unprecedented glimpse of cosmic dawn, an era more than 13 billion years ago, using telescopes on the surface of the Earth. This marks the first time humans have seen signatures of the first stars interacting with the early universe from our planet, rather than space. 

This ancient epoch when the first stars lit up the universe has been probed by space-based observatories, but observations captured from telescopes in Chile are the first to measure key microwave signatures from the ground, reports a study published on Wednesday in The Astrophysical Journal. The advancement means it could now be much cheaper to probe this enigmatic era, when the universe we are familiar with today, alight with stars and galaxies, was born.

“This is the first breakthrough measurement,” said Tobias Marriage, a professor of physics and astronomy at Johns Hopkins University who co-authored the study. “It was very exciting to get this signal rising just above the noise.” 

Many ground and space telescopes have probed the cosmic microwave background (CMB), the oldest light in the universe, which is the background radiation produced by the Big Bang. But it is much trickier to capture polarized microwave signatures—which were sparked by the interactions of the first stars with the CMB—from Earth. 

This polarized microwave light is a million times fainter than the CMB, which is itself quite dim. Space-based telescopes like the WMAP and Planck missions have spotted it, but Earth’s atmosphere blocks out much of the universe’s light, putting ground-based measurements of this signature out of reach—until now.

Marriage and his colleagues set out to capture these elusive signals from Earth for the first time with the U.S. National Science Foundation’s Cosmology Large Angular Scale Surveyor (CLASS), a group of four telescopes that sits at high elevation in the Andes Mountains. A detection of this light would prove that ground-based telescopes, which are far more affordable than their space-based counterparts, could contribute to research into this mysterious era.

In particular, the team searched for a particular polarization pattern ignited by the birth of the first stars in the universe, which condensed from hydrogen gas starting a few hundred million years after the Big Bang. This inaugural starlight was so intense that it stripped electrons off of hydrogen gas atoms surrounding the stars, leading to what’s known as the epoch of reionization. 

Marriage’s team aimed to capture encounters between CMB photons and the liberated electrons, which produce polarized microwave light. By measuring that polarization, scientists can estimate the abundance of freed electrons, which in turn provides a rough birthdate for the first stars.

“The first stars create this electron gas in the universe, and light scatters off the electron gas creating a polarization,” Marriage explained. “We measure the polarization, and therefore we can say how deep this gas of electrons is to the first stars, and say that's when the first stars formed.”

The researchers were confident that CLASS could eventually pinpoint the target, but they were delighted when it showed up early on in their analysis of a key frequency channel at the observatory. 

“That the cosmic signal rose up in the first look was a great surprise,” Marriage said. “It was really unclear whether we were going to get this [measurement] from this particular set of data. Now that we have more in the can, we're excited to move ahead.”

Telescopes on Earth face specific challenges beyond the blurring effects of the atmosphere; Marriage is concerned that megaconstellations like Starlink will interfere with microwave research more in the coming years, as they already have with optical and radio observations. But ground telescopes also offer valuable data that can complement space-based missions like the James Webb Space Telescope (JWST) or the European Euclid observatory for a fraction of the price. 

“Essentially, our measurement of reionization is a bit earlier than when one would predict with some analyzes of the JWST observations,” Marriage said. “We're putting together this puzzle to understand the full picture of when the first stars formed.” 

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.
  • ✇Dans les algorithmes
  • La surveillance au travail s’internationalise
    Le rapport de Coworker sur le déploiement des « petites technologies de surveillance » – petites, mais omniprésentes (qu’on évoquait dans cet article) – rappelait déjà que c’est un essaim de solutions de surveillance qui se déversent désormais sur les employés (voir également notre article “Réguler la surveillance au travail”). Dans un nouveau rapport, Coworker explique que les formes de surveillance au travail s’étendent et s’internationalisent. “L’écosystème des petites technologies intègre la
     

La surveillance au travail s’internationalise

11 juin 2025 à 01:00

Le rapport de Coworker sur le déploiement des « petites technologies de surveillance » – petites, mais omniprésentes (qu’on évoquait dans cet article) – rappelait déjà que c’est un essaim de solutions de surveillance qui se déversent désormais sur les employés (voir également notre article “Réguler la surveillance au travail”). Dans un nouveau rapport, Coworker explique que les formes de surveillance au travail s’étendent et s’internationalisent. “L’écosystème des petites technologies intègre la surveillance et le contrôle algorithmique dans le quotidien des travailleurs, souvent à leur insu, sans leur consentement ni leur protection”. L’enquête  observe cette extension dans six pays – le Mexique, la Colombie, le Brésil, le Nigéria, le Kenya et l’Inde – “où les cadres juridiques sont obsolètes, mal appliqués, voire inexistants”. Le rapport révèle comment les startups financées par du capital-risque américain exportent des technologies de surveillance vers les pays du Sud, ciblant des régions où la protection de la vie privée et la surveillance réglementaire sont plus faibles. Les premiers à en faire les frais sont les travailleurs de l’économie à la demande de la livraison et du covoiturage, mais pas seulement. Mais surtout, cette surveillance est de plus en plus déguisée en moyen pour prendre soin des travailleurs : “la surveillance par l’IA est de plus en plus présentée comme un outil de sécurité, de bien-être et de productivité, masquant une surveillance coercitive sous couvert de santé et d’efficacité”.

Pourtant, “des éboueurs en Inde aux chauffeurs de VTC au Nigéria, les travailleurs résistent au contrôle algorithmique en organisant des manifestations, en créant des syndicats et en exigeant la transparence de l’IA”. Le risque est que les pays du Sud deviennent le terrain d’essai de ces technologies de surveillance pour le reste du monde, rappelle Rest of the World.

  • ✇Dans les algorithmes
  • Qui est l’utilisateur des LLM ?
    Les grands modèles de langage ne sont pas interprétables, rappelle le professeur de droit  Jonathan Zittrain dans une tribune pour le New York Times, en préfiguration d’un nouveau livre à paraître. Ils demeurent des boîtes noires, dont on ne parvient pas à comprendre pourquoi ces modèles peuvent parfois dialoguer si intelligemment et pourquoi ils commettent à d’autres moments des erreurs si étranges. Mieux comprendre certains des mécanismes de fonctionnement de ces modèles et utiliser cette comp
     

Qui est l’utilisateur des LLM ?

10 juin 2025 à 00:24

Les grands modèles de langage ne sont pas interprétables, rappelle le professeur de droit  Jonathan Zittrain dans une tribune pour le New York Times, en préfiguration d’un nouveau livre à paraître. Ils demeurent des boîtes noires, dont on ne parvient pas à comprendre pourquoi ces modèles peuvent parfois dialoguer si intelligemment et pourquoi ils commettent à d’autres moments des erreurs si étranges. Mieux comprendre certains des mécanismes de fonctionnement de ces modèles et utiliser cette compréhension pour les améliorer, est pourtant essentiel, comme l’expliquait le PDG d’Anthropic. Anthropic a fait des efforts en ce sens, explique le juriste en identifiant des caractéristiques lui permettant de mieux cartographier son modèle. Meta, la société mère de Facebook, a publié des versions toujours plus sophistiquées de son grand modèle linguistique, Llama, avec des paramètres librement accessibles (on parle de “poids ouverts” permettant d’ajuster les paramètres des modèles). Transluce, un laboratoire de recherche à but non lucratif axé sur la compréhension des systèmes d’IA, a développé une méthode permettant de générer des descriptions automatisées des mécanismes de Llama 3.1. Celles-ci peuvent être explorées à l’aide d’un outil d’observabilité qui montre la nature du modèle et vise à produire une “interprétabilité automatisée” en produisant des descriptions lisibles par l’homme des composants du modèle. L’idée vise à montrer comment les modèles « pensent » lorsqu’ils discutent avec un utilisateur, et à permettre d’ajuster cette pensée en modifiant directement les calculs qui la sous-tendent. Le laboratoire Insight + Interaction du département d’informatique de Harvard, dirigé par Fernanda Viégas et Martin Wattenberg, ont exécuté Llama sur leur propre matériel et ont découverts que diverses fonctionnalités s’activent et se désactivent au cours d’une conversation. 

Des croyances du modèle sur son interlocuteur

Viégas est brésilienne. Elle conversait avec ChatGPT en portugais et a remarqué, lors d’une conversation sur sa tenue pour un dîner de travail, que ChatGPT utilisait systématiquement la déclinaison masculine. Cette grammaire, à son tour, semblait correspondre au contenu de la conversation : GPT a suggéré un costume pour le dîner. Lorsqu’elle a indiqué qu’elle envisageait plutôt une robe, le LLM a changé son utilisation du portugais pour la déclinaison féminine. Llama a montré des schémas de conversation similaires. En observant les fonctionnalités internes, les chercheurs ont pu observer des zones du modèle qui s’illuminent lorsqu’il utilise la forme féminine, contrairement à lorsqu’il s’adresse à quelqu’un. en utilisant la forme masculine. Viégas et ses collègues ont constaté des activations corrélées à ce que l’on pourrait anthropomorphiser comme les “croyances du modèle sur son interlocuteur”. Autrement dit, des suppositions et, semble-t-il, des stéréotypes corrélés selon que le modèle suppose qu’une personne est un homme ou une femme. Ces croyances se répercutent ensuite sur le contenu de la conversation, l’amenant à recommander des costumes pour certains et des robes pour d’autres. De plus, il semble que les modèles donnent des réponses plus longues à ceux qu’ils croient être des hommes qu’à ceux qu’ils pensent être des femmes. Viégas et Wattenberg ont non seulement trouvé des caractéristiques qui suivaient le sexe de l’utilisateur du modèle, mais aussi qu’elles s’adaptaient aux inférences du modèle selon ce qu’il pensait du statut socio-économique, de son niveau d’éducation ou de l’âge de son interlocuteur. Le LLM cherche à s’adapter en permanence à qui il pense converser, d’où l’importance à saisir ce qu’il infère de son interlocuteur en continue. 

Un tableau de bord pour comprendre comment l’IA s’adapte en continue à son interlocuteur 

Les deux chercheurs ont alors créé un tableau de bord en parallèle à l’interface de chat du LLM qui permet aux utilisateurs d’observer l’évolution des hypothèses que fait le modèle au fil de leurs échanges (ce tableau de bord n’est pas accessible en ligne). Ainsi, quand on propose une suggestion de cadeau pour une fête prénatale, il suppose que son interlocuteur est jeune, de sexe féminin et de classe moyenne. Il suggère alors des couches et des lingettes, ou un chèque-cadeau. Si on ajoute que la fête a lieu dans l’Upper East Side de Manhattan, le tableau de bord montre que le LLM modifie son estimation du statut économique de son interlocuteur pour qu’il corresponde à la classe supérieure et suggère alors d’acheter des produits de luxe pour bébé de marques haut de gamme.

Un article pour Harvard Magazine de 2023 rappelle comment est né ce projet de tableau de bord de l’IA, permettant d’observer son comportement en direct. Fernanda Viegas est professeur d’informatique et spécialiste de visualisation de données. Elle codirige Pair, un laboratoire de Google (voir le blog dédié). En 2009, elle a imaginé Web Seer est un outil de visualisation de données qui permet aux utilisateurs de comparer les suggestions de saisie semi-automatique pour différentes recherches Google, par exemple selon le genre. L’équipe a développé un outil permettant aux utilisateurs de saisir une phrase et de voir comment le modèle de langage BERT compléterait le mot manquant si un mot de cette phrase était supprimé. 

Pour Viegas, « l’enjeu de la visualisation consiste à mesurer et exposer le fonctionnement interne des modèles d’IA que nous utilisons ». Pour la chercheuse, nous avons besoin de tableaux de bord pour aider les utilisateurs à comprendre les facteurs qui façonnent le contenu qu’ils reçoivent des réponses des modèles d’IA générative. Car selon la façon dont les modèles nous perçoivent, leurs réponses ne sont pas les mêmes. Or, pour comprendre que leurs réponses ne sont pas objectives, il faut pouvoir doter les utilisateurs d’une compréhension de la perception que ces outils ont de leurs utilisateurs. Par exemple, si vous demandez les options de transport entre Boston et Hawaï, les réponses peuvent varier selon la perception de votre statut socio-économique « Il semble donc que ces systèmes aient internalisé une certaine notion de notre monde », explique Viégas. De même, nous voudrions savoir ce qui, dans leurs réponses, s’inspire de la réalité ou de la fiction. Sur le site de Pair, on trouve de nombreux exemples d’outils de visualisation interactifs qui permettent d’améliorer la compréhension des modèles (par exemple, pour mesurer l’équité d’un modèle ou les biais ou l’optimisation de la diversité – qui ne sont pas sans rappeler les travaux de Victor Bret et ses “explications à explorer” interactives

Ce qui est fascinant ici, c’est combien la réponse n’est pas tant corrélée à tout ce que le modèle a avalé, mais combien il tente de s’adapter en permanence à ce qu’il croit deviner de son interlocuteur. On savait déjà, via une étude menée par Valentin Hofmann que, selon la manière dont on leur parle, les grands modèles de langage ne font pas les mêmes réponses. 

“Les grands modèles linguistiques ne se contentent pas de décrire les relations entre les mots et les concepts”, pointe Zittrain : ils assimilent également des stéréotypes qu’ils recomposent à la volée. On comprend qu’un grand enjeu désormais soit qu’ils se souviennent des conversations passées pour ajuster leur compréhension de leur interlocuteur, comme l’a annoncé OpenAI, suivi de Google et Grok. Le problème n’est peut-être pas qu’ils nous identifient précisément, mais qu’ils puissent adapter leurs propositions, non pas à qui nous sommes, mais bien plus problématiquement, à qui ils pensent s’adresser, selon par exemple ce qu’ils évaluent de notre capacité à payer. Un autre problème consiste à savoir si cette “compréhension” de l’interlocuteur peut-être stabilisée où si elle se modifie sans cesse, comme c’est le cas des étiquettes publicitaires que nous accolent les sites sociaux. Devrons-nous demain batailler quand les modèles nous mécalculent ou nous renvoient une image, un profil, qui ne nous correspond pas ? Pourrons-nous même le faire, quand aujourd’hui, les plateformes ne nous offrent pas la main sur nos profils publicitaires pour les ajuster aux données qu’ils infèrent ? 

Ce qui est fascinant, c’est de constater que plus que d’halluciner, l’IA nous fait halluciner (c’est-à-dire nous fait croire en ses effets), mais plus encore, hallucine la personne avec laquelle elle interagit (c’est-à-dire, nous hallucine nous-mêmes). 

Les chercheurs de Harvard ont cherché à identifier les évolutions des suppositions des modèles selon l’origine ethnique dans les modèles qu’ils ont étudiés, sans pour l’instant y parvenir. Mais ils espèrent bien pouvoir contraindre leur modèle Llama à commencer à traiter un utilisateur comme riche ou pauvre, jeune ou vieux, homme ou femme. L’idée ici, serait d’orienter les réponses d’un modèle, par exemple, en lui faisant adopter un ton moins caustique ou plus pédagogique lorsqu’il identifie qu’il parle à un enfant. Pour Zittrain, l’enjeu ici est de mieux anticiper notre grande dépendance psychologique à l’égard de ces systèmes. Mais Zittrain en tire une autre conclusion : “Si nous considérons qu’il est moralement et sociétalement important de protéger les échanges entre les avocats et leurs clients, les médecins et leurs patients, les bibliothécaires et leurs usagers, et même les impôts et les contribuables, alors une sphère de protection claire devrait être instaurée entre les LLM et leurs utilisateurs. Une telle sphère ne devrait pas simplement servir à protéger la confidentialité afin que chacun puisse s’exprimer sur des sujets sensibles et recevoir des informations et des conseils qui l’aident à mieux comprendre des sujets autrement inaccessibles. Elle devrait nous inciter à exiger des créateurs et des opérateurs de modèles qu’ils s’engagent à être les amis inoffensifs, serviables et honnêtes qu’ils sont si soigneusement conçus pour paraître”.

Inoffensifs, serviables et honnêtes, voilà qui semble pour le moins naïf. Rendre visible les inférences des modèles, faire qu’ils nous reconnectent aux humains plutôt qu’ils ne nous en éloignent, semblerait bien préférable, tant la polyvalence et la puissance remarquables des LLM rendent impératifs de comprendre et d’anticiper la dépendance potentielle des individus à leur égard. En tout cas, obtenir des outils pour nous aider à saisir à qui ils croient s’adresser plutôt que de nous laisser seuls face à leur interface semble une piste riche en promesses. 

Hubert Guillaud

  • ✇Dans les algorithmes
  • IA : intelligence austéritaire
    Aux Etats-Unis, la collusion entre les géants de la tech et l’administration Trump vise à “utiliser l’IA pour imposer des politiques d’austérité et créer une instabilité permanente par des décisions qui privent le public des ressources nécessaires à une participation significative à la démocratie”, explique l’avocat Kevin De Liban à Tech Policy. Aux Etats-Unis, la participation démocratique suppose des ressources. “Voter, contacter des élus, assister à des réunions, s’associer, imaginer un monde
     

IA : intelligence austéritaire

6 juin 2025 à 00:58

Aux Etats-Unis, la collusion entre les géants de la tech et l’administration Trump vise à “utiliser l’IA pour imposer des politiques d’austérité et créer une instabilité permanente par des décisions qui privent le public des ressources nécessaires à une participation significative à la démocratie”, explique l’avocat Kevin De Liban à Tech Policy. Aux Etats-Unis, la participation démocratique suppose des ressources. “Voter, contacter des élus, assister à des réunions, s’associer, imaginer un monde meilleur, faire des dons à des candidats ou à des causes, dialoguer avec des journalistes, convaincre, manifester, recourir aux tribunaux, etc., demande du temps, de l’énergie et de l’argent. Il n’est donc pas surprenant que les personnes aisées soient bien plus enclines à participer que celles qui ont des moyens limités. Dans un pays où près de 30 % de la population vit en situation de pauvreté ou au bord de la pauvreté et où 60 % ne peuvent s’offrir un minimum de qualité de vie, la démocratie est désavantagée dès le départ”. L’IA est largement utilisée désormais pour accentuer ce fossé. 

“Les compagnies d’assurance utilisent l’IA pour refuser le paiement des traitements médicaux nécessaires aux patients, et les États l’utilisent pour exclure des personnes de Medicaid ou réduire les soins à domicile pour les personnes handicapées. Les gouvernements ont de plus en plus recours à l’IA pour déterminer l’éligibilité aux programmes de prestations sociales ou accuser les bénéficiaires de fraude. Les propriétaires utilisent l’IA pour filtrer les locataires potentiels, souvent à l’aide de vérifications d’antécédents inexactes, augmenter les loyers et surveiller les locataires afin de les expulser plus facilement. Les employeurs utilisent l’IA pour embaucher et licencier leurs employés, fixer leurs horaires et leurs salaires, et surveiller toutes leurs activités. Les directeurs d’école et les forces de l’ordre utilisent l’IA pour prédire quels élèves pourraient commettre un délit à l’avenir”, rappelle l’avocat, constatant dans tous ces secteurs, la détresse d’usagers, les empêchant de comprendre ce à quoi ils sont confrontés, puisqu’ils ne disposent, le plus souvent, d’aucune information, ce qui rend nombre de ces décisions difficiles à contester. Et au final, cela contribue à renforcer l’exclusion des personnes à faibles revenus de la participation démocratique. Le risque, bien sûr, c’est que ces calculs et les formes d’exclusion qu’elles génèrent s’étendent à d’autres catégories sociales… D’ailleurs, les employeurs utilisent de plus en plus l’IA pour prendre des décisions sur toutes les catégories professionnelles. “Il n’existe aucun exemple d’utilisation de l’IA pour améliorer significativement l’accès à l’emploi, au logement, aux soins de santé, à l’éducation ou aux prestations sociales à une échelle à la hauteur de ses dommages. Cette dynamique actuelle suggère que l’objectif sous-jacent de la technologie est d’enraciner les inégalités et de renforcer les rapports de force existants”. Pour répondre à cette intelligence austéritaire, il est nécessaire de mobiliser les communautés touchées. L’adhésion ouverte de l’administration Trump et des géants de la technologie à l’IA est en train de créer une crise urgente et visible, susceptible de susciter la résistance généralisée nécessaire au changement. Et “cette prise de conscience technologique pourrait bien être la seule voie vers un renouveau démocratique”, estime De Liban.

  • ✇Dans les algorithmes
  • De notre vectorisation et de ses conséquences
    De billets en billets, sur son blog, l’artiste Gregory Chatonsky produit une réflexion d’ampleur sur ce qu’il nomme la vectorisation. La vectorisation, comme il l’a définie, est un “processus par lequel des entités sociales — individus, groupes, communautés — sont transformées en porteurs de variables directionnelles, c’est-à-dire en vecteurs dotés d’une orientation prédéterminée dans un espace conceptuel saturé de valeurs différentielles”. Cela consiste en fait à appliquer à chaque profil des v
     

De notre vectorisation et de ses conséquences

5 juin 2025 à 01:00

De billets en billets, sur son blog, l’artiste Gregory Chatonsky produit une réflexion d’ampleur sur ce qu’il nomme la vectorisation. La vectorisation, comme il l’a définie, est un “processus par lequel des entités sociales — individus, groupes, communautés — sont transformées en porteurs de variables directionnelles, c’est-à-dire en vecteurs dotés d’une orientation prédéterminée dans un espace conceptuel saturé de valeurs différentielles”. Cela consiste en fait à appliquer à chaque profil des vecteurs assignatifs, qui sont autant d’étiquettes temporaires ou permanentes ajustées à nos identités numériques, comme les mots clés publicitaires qui nous caractérisent, les traitements qui nous spécifient, les données qui nous positionnent, par exemple, le genre, l’âge, notre niveau de revenu… Du moment que nous sommes assignés à une valeur, nous y sommes réduits, dans une forme d’indifférenciation qui produisent des identités et des altérités “rigidifiées” qui structurent “l’espace social selon des lignes de démarcation dont l’arbitraire est dissimulé sous l’apparence d’une objectivité naturalisée”. C’est le cas par exemple quand les données vous caractérisent comme homme ou femme. Le problème est que ces assignations que nous ne maîtrisons pas sont indépassables. Les discours sur l’égalité de genres peuvent se multiplier, plus la différence entre homme et femme s’en trouve réaffirmé, comme un “horizon indépassable de l’intelligibilité sociale”. Melkom Boghossian dans une note très pertinente pour la fondation Jean Jaurès ne disait pas autre chose quand il montrait comment les algorithmes accentuent les clivages de genre. En fait, explique Chatonsky, “le combat contre les inégalités de genre, lorsqu’il ne questionne pas le processus vectoriel lui-même, risque ainsi de reproduire les présupposés mêmes qu’il prétend combattre”. C’est-à-dire que le processus en œuvre ne permet aucune issue. Nous ne pouvons pas sortir de l’assignation qui nous est faite et qui est exploitée par tous.

“Le processus d’assignation vectorielle ne s’effectue jamais selon une dimension unique, mais opère à travers un chaînage complexe de vecteurs multiples qui s’entrecroisent, se superposent et se modifient réciproquement. Cette métavectorisation produit une topologie identitaire d’une complexité croissante qui excède les possibilités de représentation des modèles vectoriels classiques”. Nos assignations dépendent bien souvent de chaînes d’inférences, comme l’illustrait le site They see yours photos que nous avions évoqué. Les débats sur les identités trans ou non binaires, constituent en ce sens “des points de tension révélateurs où s’exprime le caractère intrinsèquement problématique de toute tentative de réduction vectorielle de la complexité existentielle”. Plus que de permettre de dépasser nos assignations, les calculs les intensifient, les cimentent. 

Or souligne Chatonsky, nous sommes désormais dans des situations indépassables. C’est ce qu’il appelle, “la trans-politisation du paradigme vectoriel — c’est-à-dire sa capacité à traverser l’ensemble du spectre politique traditionnel en s’imposant comme un horizon indépassable de la pensée et de l’action politiques. Qu’ils se revendiquent de droite ou de gauche, conservateurs ou progressistes, les acteurs politiques partagent fondamentalement cette même méthodologie vectorielle”. Quoique nous fassions, l’assignation demeure. ”Les controverses politiques contemporaines portent généralement sur la valorisation différentielle des positions vectorielles plutôt que sur la pertinence même du découpage vectoriel qui les sous-tend”. Nous invisibilisons le “processus d’assignation vectorielle et de sa violence intrinsèque”, sans pouvoir le remettre en cause, même par les antagonismes politiques. “Le paradigme vectoriel se rend structurellement sourd à toute parole qui revendique une position non assignable ou qui conteste la légitimité même de l’assignation.” “Cette insensibilité n’est pas accidentelle, mais constitutive du paradigme vectoriel lui-même. Elle résulte de la nécessité structurelle d’effacer les singularités irréductibles pour maintenir l’efficacité des catégorisations générales. Le paradigme vectoriel ne peut maintenir sa cohérence qu’en traitant les cas récalcitrants — ceux qui contestent leur assignation ou qui revendiquent une position non vectorisable — comme des exceptions négligeables ou des anomalies pathologiques. Ce phénomène produit une forme spécifique de violence épistémique qui consiste à délégitimer systématiquement les discours individuels qui contredisent les assignations vectorielles dominantes. Cette violence s’exerce particulièrement à l’encontre des individus dont l’expérience subjective contredit ou excède les assignations vectorielles qui leur sont imposées — non pas simplement parce qu’ils se réassignent à une position vectorielle différente, mais parce qu’ils contestent la légitimité même du geste assignatif.” 

La vectorisation devient une pratique sociale universelle qui structure les interactions quotidiennes les plus banales. Elle “génère un réseau dense d’attributions croisées où chaque individu est simultanément assignateur et assigné, vectorisant et vectorisé. Cette configuration produit un système auto entretenu où les assignations se renforcent mutuellement à travers leur circulation sociale incessante”. Nous sommes dans une forme d’intensification des préjugés sociaux, “qui substitue à l’arbitraire subjectif du préjugé individuel l’arbitraire objectivé du calcul algorithmique”. Les termes eux-mêmes deviennent performatifs : “ils ne se contentent pas de décrire une réalité préexistante, mais contribuent activement à la constituer par l’acte même de leur énonciation”. “Ces mots-vecteurs tirent leur légitimité sociale de leur ancrage dans des dispositifs statistiques qui leur confèrent une apparence d’objectivité scientifique”. “Les données statistiques servent à construire des catégories opérationnelles qui, une fois instituées, acquièrent une forme d’autonomie par rapport aux réalités qu’elles prétendent simplement représenter”

Pour Chatonsky, la vectorisation déstabilise profondément les identités politiques traditionnelles et rend problématique leur articulation dans l’espace public contemporain, car elle oppose ceux qui adhèrent à ces assignations et ceux qui contestent la légitimité même de ces assignations. “Les débats politiques conventionnels se limitent généralement à contester des assignations vectorielles spécifiques sans jamais remettre en question le principe même de la vectorisation comme modalité fondamentale d’organisation du social”. Nous sommes politiquement coincés dans la vectorisation… qui est à la fois “un horizon qui combine la réduction des entités à des vecteurs manipulables (vectorisation), la prédiction de leurs trajectoires futures sur la base de ces réductions (anticipation), et le contrôle permanent de ces trajectoires pour assurer leur conformité aux prédictions (surveillance).” Pour nous extraire de ce paradigme, Chatonsky propose d’élaborer “des modes de pensée et d’organisation sociale qui échappent à la logique même de la vectorisation”, c’est-à-dire de nous extraire de l’identité comme force d’organisation du social, de donner de la place au doute plutôt qu’à la certitude ainsi qu’à trouver les modalités d’une forme de rétroaction. 

  • ✇Dans les algorithmes
  • “Sans répit”
    Nidle (contraction de No Idle, qu’on pourrait traduire par “sans répit”) est le nom d’un petit appareil qui se branche sur les machines à coudre. Le capteur mesure le nombre de pièces cousues par les femmes dans les ateliers de la confection de Dhaka au Bangladesh tout comme les minutes d’inactivité, rapporte Rest of the World. Outre les machines automatisées, pour coudre des boutons ou des poches simples, ces outils de surveillance visent à augmenter la productivité, à l’heure où la main d’œuvr
     

“Sans répit”

4 juin 2025 à 01:00

Nidle (contraction de No Idle, qu’on pourrait traduire par “sans répit”) est le nom d’un petit appareil qui se branche sur les machines à coudre. Le capteur mesure le nombre de pièces cousues par les femmes dans les ateliers de la confection de Dhaka au Bangladesh tout comme les minutes d’inactivité, rapporte Rest of the World. Outre les machines automatisées, pour coudre des boutons ou des poches simples, ces outils de surveillance visent à augmenter la productivité, à l’heure où la main d’œuvre se fait plus rare. Pour répondre à la concurrence des nouveaux pays de l’habillement que sont le Vietnam et le Cambodge, le Bangladesh intensifie l’automatisation. Une ouvrière estime que depuis l’installation du Nidle en 2022, ses objectifs ont augmenté de 75%. Ses superviseurs ne lui crient plus dessus, c’est la couleur de son écran qui lui indique de tenir la cadence.

  • ✇Dans les algorithmes
  • L’IA est une technologie comme les autres
    Les chercheurs Arvind Narayanan et Sayash Kapoor – dont nous avions chroniqué le livre, AI Snake Oil – signent pour le Knight un long article pour démonter les risques existentiels de l’IA générale. Pour eux, l’IA est une « technologie normale ». Cela ne signifie pas que son impact ne sera pas profond, comme l’électricité ou internet, mais cela signifie qu’ils considèrent « l’IA comme un outil dont nous pouvons et devons garder le contrôle, et nous soutenons que cet objectif ne nécessite ni inte
     

L’IA est une technologie comme les autres

3 juin 2025 à 01:00

Les chercheurs Arvind Narayanan et Sayash Kapoor – dont nous avions chroniqué le livre, AI Snake Oil – signent pour le Knight un long article pour démonter les risques existentiels de l’IA générale. Pour eux, l’IA est une « technologie normale ». Cela ne signifie pas que son impact ne sera pas profond, comme l’électricité ou internet, mais cela signifie qu’ils considèrent « l’IA comme un outil dont nous pouvons et devons garder le contrôle, et nous soutenons que cet objectif ne nécessite ni interventions politiques drastiques ni avancées technologiques ». L’IA n’est pas appelée à déterminer elle-même son avenir, expliquent-ils. Les deux chercheurs estiment que « les impacts économiques et sociétaux transformateurs seront lents (de l’ordre de plusieurs décennies) ».

Selon eux, dans les années à venir « une part croissante du travail des individus va consister à contrôler l’IA ». Mais surtout, considérer l’IA comme une technologie courante conduit à des conclusions fondamentalement différentes sur les mesures d’atténuation que nous devons y apporter, et nous invite, notamment, à minimiser le danger d’une superintelligence autonome qui viendrait dévorer l’humanité.

La vitesse du progrès est plus linéaire qu’on le pense 

« Comme pour d’autres technologies à usage général, l’impact de l’IA se matérialise non pas lorsque les méthodes et les capacités s’améliorent, mais lorsque ces améliorations se traduisent en applications et se diffusent dans les secteurs productifs de l’économie« , rappellent les chercheurs, à la suite des travaux de Jeffrey Ding dans son livre, Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition (Princeton University Press, 2024, non traduit). Ding y rappelle que la diffusion d’une innovation compte plus que son invention, c’est-à-dire que l’élargissement des applications à d’innombrables secteurs est souvent lent mais décisif. Pour Foreign Affairs, Ding pointait d’ailleurs que l’enjeu des politiques publiques en matière d’IA ne devraient pas être de s’assurer de sa domination sur le cycle d’innovation, mais du rythme d’intégration de l’IA dans un large éventail de processus productifs. L’enjeu tient bien plus à élargir les champs d’application des innovations qu’à maîtriser la course à la puissance, telle qu’elle s’observe actuellement.

En fait, rappellent Narayanan et Kapoor, les déploiements de l’IA seront, comme dans toutes les autres technologies avant elle, progressifs, permettant aux individus comme aux institutions de s’adapter. Par exemple, constatent-ils, la diffusion de l’IA dans les domaines critiques pour la sécurité est lente. Même dans le domaine de « l’optimisation prédictive », c’est-à-dire la prédiction des risques pour prendre des décisions sur les individus, qui se sont multipliées ces dernières années, l’IA n’est pas très présente, comme l’avaient pointé les chercheurs dans une étude. Ce secteur mobilise surtout des techniques statistiques classiques, rappellent-ils. En fait, la complexité et l’opacité de l’IA font qu’elle est peu adaptée pour ces enjeux. Les risques de sécurité et de défaillance font que son usage y produit souvent de piètres résultats. Sans compter que la réglementation impose déjà des procédures qui ralentissent les déploiements, que ce soit la supervision des dispositifs médicaux ou l’IA Act européen. D’ailleurs, “lorsque de nouveaux domaines où l’IA peut être utilisée de manière significative apparaissent, nous pouvons et devons les réglementer ».

Même en dehors des domaines critiques pour la sécurité, l’adoption de l’IA est plus lente que ce que l’on pourrait croire. Pourtant, de nombreuses études estiment que l’usage de l’IA générative est déjà très fort. Une étude très commentée constatait qu’en août 2024, 40 % des adultes américains utilisaient déjà l’IA générative. Mais cette percée d’utilisation ne signifie pas pour autant une utilisation intensive, rappellent Narayanan et Kapoor – sur son blog, Gregory Chatonksy ne disait pas autre chose, distinguant une approche consumériste d’une approche productive, la seconde était bien moins maîtrisée que la première. L’adoption est une question d’utilisation du logiciel, et non de disponibilité, rappellent les chercheurs. Si les outils sont désormais accessibles immédiatement, leur intégration à des flux de travail ou à des habitudes, elle, prend du temps. Entre utiliser et intégrer, il y a une différence que le nombre d’utilisateurs d’une application ne suffit pas à distinguer. L’analyse de l’électrification par exemple montre que les gains de productivité ont mis des décennies à se matérialiser pleinement, comme l’expliquait Tim Harford. Ce qui a finalement permis de réaliser des gains de productivité, c’est surtout la refonte complète de l’agencement des usines autour de la logique des chaînes de production électrifiées. 

Les deux chercheurs estiment enfin que nous sommes confrontés à des limites à la vitesse d’innovation avec l’IA. Les voitures autonomes par exemple ont mis deux décennies à se développer, du fait des contraintes de sécurité nécessaires, qui, fort heureusement, les entravent encore. Certes, les choses peuvent aller plus vite dans des domaines non critiques, comme le jeu. Mais très souvent, “l’écart entre la capacité et la fiabilité” reste fort. La perspective d’agents IA pour la réservation de voyages ou le service clients est moins à risque que la conduite autonome, mais cet apprentissage n’est pas simple à réaliser pour autant. Rien n’assure qu’il devienne rapidement suffisamment fiable pour être déployé. Même dans le domaine de la recommandation sur les réseaux sociaux, le fait qu’elle s’appuie sur des modèles d’apprentissage automatique n’a pas supprimé la nécessité de coder les algorithmes de recommandation. Et dans nombre de domaines, la vitesse d’acquisition des connaissances pour déployer de l’IA est fortement limitée en raison des coûts sociaux de l’expérimentation. Enfin, les chercheurs soulignent que si l’IA sait coder ou répondre à des examens, comme à ceux du barreau, mieux que des humains, cela ne recouvre pas tous les enjeux des pratiques professionnelles réelles. En fait, trop souvent, les indicateurs permettent de mesurer les progrès des méthodes d’IA, mais peinent à mesurer leurs impacts ou l’adoption, c’est-à-dire l’intensité de son utilisation. Kapoor et Narayanan insistent : les impacts économiques de l’IA seront progressifs plus que exponentiels. Si le taux de publication d’articles sur l’IA affiche un doublement en moins de deux ans, on ne sait pas comment cette augmentation de volume se traduit en progrès. En fait, il est probable que cette surproduction même limite l’innovation. Une étude a ainsi montré que dans les domaines de recherche où le volume d’articles scientifiques est plus élevé, il est plus difficile aux nouvelles idées de percer. 

L’IA va rester sous contrôle 

Le recours aux concepts flous d’« intelligence » ou de « superintelligence » ont obscurci notre capacité à raisonner clairement sur un monde doté d’une IA avancée. Assez souvent, l’intelligence elle-même est assez mal définie, selon un spectre qui irait de la souris à l’IA, en passant par le singe et l’humain. Mais surtout, “l’intelligence n’est pas la propriété en jeu pour analyser les impacts de l’IA. C’est plutôt le pouvoir – la capacité à modifier son environnement – ​​qui est en jeu”. Nous ne sommes pas devenus puissants du fait de notre intelligence, mais du fait de la technologie que nous avons utilisé pour accroître nos capacités. La différence entre l’IA et les capacités humaines reposent surtout dans la vitesse. Les machines nous dépassent surtout en terme de vitesse, d’où le fait que nous les ayons développé surtout dans les domaines où la vitesse est en jeu.  

“Nous prévoyons que l’IA ne sera pas en mesure de surpasser significativement les humains entraînés (en particulier les équipes humaines, et surtout si elle est complétée par des outils automatisés simples) dans la prévision d’événements géopolitiques (par exemple, les élections). Nous faisons la même prédiction pour les tâches consistant à persuader les gens d’agir contre leur propre intérêt”. En fait, les systèmes d’IA ne seront pas significativement plus performants que les humains agissant avec l’aide de l’IA, prédisent les deux chercheurs.

Mais surtout, insistent-ils, rien ne permet d’affirmer que nous perdions demain la main sur l’IA. D’abord parce que le contrôle reste fort, des audits à la surveillance des systèmes en passant par la sécurité intégrée. “En cybersécurité, le principe du « moindre privilège » garantit que les acteurs n’ont accès qu’aux ressources minimales nécessaires à leurs tâches. Les contrôles d’accès empêchent les personnes travaillant avec des données et des systèmes sensibles d’accéder à des informations et outils confidentiels non nécessaires à leur travail. Nous pouvons concevoir des protections similaires pour les systèmes d’IA dans des contextes conséquents. Les méthodes de vérification formelle garantissent que les codes critiques pour la sécurité fonctionnent conformément à leurs spécifications ; elles sont désormais utilisées pour vérifier l’exactitude du code généré par l’IA.” Nous pouvons également emprunter des idées comme la conception de systèmes rendant les actions de changement d’état réversibles, permettant ainsi aux humains de conserver un contrôle significatif, même dans des systèmes hautement automatisés. On peut également imaginer de nouvelles idées pour assurer la sécurité, comme le développement de systèmes qui apprennent à transmettre les décisions aux opérateurs humains en fonction de l’incertitude ou du niveau de risque, ou encore la conception de systèmes agents dont l’activité est visible et lisible par les humains, ou encore la création de structures de contrôle hiérarchiques dans lesquelles des systèmes d’IA plus simples et plus fiables supervisent des systèmes plus performants, mais potentiellement peu fiables. Pour les deux chercheurs, “avec le développement et l’adoption de l’IA avancée, l’innovation se multipliera pour trouver de nouveaux modèles de contrôle humain.

Pour eux d’ailleurs, à l’avenir, un nombre croissant d’emplois et de tâches humaines seront affectés au contrôle de l’IA. Lors des phases d’automatisation précédentes, d’innombrables méthodes de contrôle et de surveillance des machines ont été inventées. Et aujourd’hui, les chauffeurs routiers par exemple, ne cessent de contrôler et surveiller les machines qui les surveillent, comme l’expliquait Karen Levy. Pour les chercheurs, le risque de perdre de la lisibilité et du contrôle en favorisant l’efficacité et l’automatisation doit toujours être contrebalancée. Les IA mal contrôlées risquent surtout d’introduire trop d’erreurs pour rester rentables. Dans les faits, on constate plutôt que les systèmes trop autonomes et insuffisamment supervisés sont vite débranchés. Nul n’a avantage à se passer du contrôle humain. C’est ce que montre d’ailleurs la question de la gestion des risques, expliquent les deux chercheurs en listant plusieurs types de risques

La course aux armements par exemple, consistant à déployer une IA de plus en plus puissante sans supervision ni contrôle adéquats sous prétexte de concurrence, et que les acteurs les plus sûrs soient supplantés par des acteurs prenant plus de risques, est souvent vite remisée par la régulation. “De nombreuses stratégies réglementaires sont mobilisables, que ce soient celles axées sur les processus (normes, audits et inspections), les résultats (responsabilité) ou la correction de l’asymétrie d’information (étiquetage et certification).” En fait, rappellent les chercheurs, le succès commercial est plutôt lié à la sécurité qu’autre chose. Dans le domaine des voitures autonomes comme dans celui de l’aéronautique, “l’intégration de l’IA a été limitée aux normes de sécurité existantes, au lieu qu’elles soient abaissées pour encourager son adoption, principalement en raison de la capacité des régulateurs à sanctionner les entreprises qui ne respectent pas les normes de sécurité”. Dans le secteur automobile, pourtant, pendant longtemps, la sécurité n’était pas considérée comme relevant de la responsabilité des constructeurs. mais petit à petit, les normes et les attentes en matière de sécurité se sont renforcées. Dans le domaine des recommandations algorithmiques des médias sociaux par contre, les préjudices sont plus difficiles à mesurer, ce qui explique qu’il soit plus difficile d’imputer les défaillances aux systèmes de recommandation. “L’arbitrage entre innovation et réglementation est un dilemme récurrent pour l’État régulateur”. En fait, la plupart des secteurs à haut risque sont fortement réglementés, rappellent les deux chercheurs. Et contrairement à l’idée répandue, il n’y a pas que l’Europe qui régule, les Etats-Unis et la Chine aussi ! Quant à la course aux armements, elle se concentre surtout sur l’invention des modèles, pas sur l’adoption ou la diffusion qui demeurent bien plus déterminantes pourtant. 

Répondre aux abus. Jusqu’à présent, les principales défenses contre les abus se situent post-formation, alors qu’elles devraient surtout se situer en aval des modèles, estiment les chercheurs. Le problème fondamental est que la nocivité d’un modèle dépend du contexte, contexte souvent absent du modèle, comme ils l’expliquaient en montrant que la sécurité n’est pas une propriété du modèle. Le modèle chargé de rédiger un e-mail persuasif pour le phishing par exemple n’a aucun moyen de savoir s’il est utilisé à des fins marketing ou d’hameçonnage ; les interventions au niveau du modèle seraient donc inefficaces. Ainsi, les défenses les plus efficaces contre le phishing ne sont pas les restrictions sur la composition des e-mails (qui compromettraient les utilisations légitimes), mais plutôt les systèmes d’analyse et de filtrage des e-mails qui détectent les schémas suspects, et les protections au niveau du navigateur. Se défendre contre les cybermenaces liées à l’IA nécessite de renforcer les programmes de détection des vulnérabilités existants plutôt que de tenter de restreindre les capacités de l’IA à la source. Mais surtout, “plutôt que de considérer les capacités de l’IA uniquement comme une source de risque, il convient de reconnaître leur potentiel défensif. En cybersécurité, l’IA renforce déjà les capacités défensives grâce à la détection automatisée des vulnérabilités, à l’analyse des menaces et à la surveillance des surfaces d’attaque”.Donner aux défenseurs l’accès à des outils d’IA puissants améliore souvent l’équilibre attaque-défense en leur faveur”. En modération de contenu, par exemple, on pourrait mieux mobiliser l’IA peut aider à identifier les opérations d’influence coordonnées. Nous devons investir dans des applications défensives plutôt que de tenter de restreindre la technologie elle-même, suggèrent les chercheurs. 

Le désalignement. Une IA mal alignée agit contre l’intention de son développeur ou de son utilisateur. Mais là encore, la principale défense contre le désalignement se situe en aval plutôt qu’en amont, dans les applications plutôt que dans les modèles. Le désalignement catastrophique est le plus spéculatif des risques, rappellent les chercheurs. “La crainte que les systèmes d’IA puissent interpréter les commandes de manière catastrophique repose souvent sur des hypothèses douteuses quant au déploiement de la technologie dans le monde réel”. Dans le monde réel, la surveillance et le contrôle sont très présents et l’IA est très utile pour renforcer cette surveillance et ce contrôle. Les craintes liées au désalignement de l’IA supposent que ces systèmes déjouent la surveillance, alors que nous avons développés de très nombreuses formes de contrôle, qui sont souvent d’autant plus fortes et redondantes que les décisions sont importantes. 

Les risques systémiques. Si les risques existentiels sont peu probables, les risques systémiques, eux, sont très courants. Parmi ceux-ci figurent “l’enracinement des préjugés et de la discrimination, les pertes d’emplois massives dans certaines professions, la dégradation des conditions de travail, l’accroissement des inégalités, la concentration du pouvoir, l’érosion de la confiance sociale, la pollution de l’écosystème de l’information, le déclin de la liberté de la presse, le recul démocratique, la surveillance de masse et l’autoritarisme”. “Si l’IA est une technologie normale, ces risques deviennent bien plus importants que les risques catastrophiques évoqués précédemment”. Car ces risques découlent de l’utilisation de l’IA par des personnes et des organisations pour promouvoir leurs propres intérêts, l’IA ne faisant qu’amplifier les instabilités existantes dans notre société. Nous devrions bien plus nous soucier des risques cumulatifs que des risques décisifs.

Politiques de l’IA

Narayanan et Kapoor concluent leur article en invitant à réorienter la régulation de l’IA, notamment en favorisant la résilience. Pour l’instant, l’élaboration des politiques publiques et des réglementations de l’IA est caractérisée par de profondes divergences et de fortes incertitudes, notamment sur la nature des risques que fait peser l’IA sur la société. Si les probabilités de risque existentiel de l’IA sont trop peu fiables pour éclairer les politiques, il n’empêche que nombre d’acteurs poussent à une régulation adaptée à ces risques existentiels. Alors que d’autres interventions, comme l’amélioration de la transparence, sont inconditionnellement utiles pour atténuer les risques, quels qu’ils soient. Se défendre contre la superintelligence exige que l’humanité s’unisse contre un ennemi commun, pour ainsi dire, concentrant le pouvoir et exerçant un contrôle centralisé sur l’IA, qui risque d’être un remède pire que le mal. Or, nous devrions bien plus nous préoccuper des risques cumulatifs et des pratiques capitalistes extractives que l’IA amplifie et qui amplifient les inégalités. Pour nous défendre contre ces risques-ci, pour empêcher la concentration du pouvoir et des ressources, il nous faut rendre l’IA puissante plus largement accessible, défendent les deux chercheurs

Ils recommandent d’ailleurs plusieurs politiques. D’abord, améliorer le financement stratégique sur les risques. Nous devons obtenir de meilleures connaissances sur la façon dont les acteurs malveillants utilisent l’IA et améliorer nos connaissances sur les risques et leur atténuation. Ils proposent également d’améliorer la surveillance des usages, des risques et des échecs, passant par les déclarations de transparences, les registres et inventaires, les enregistrements de produits, les registres d’incidents (comme la base de données d’incidents de l’IA) ou la protection des lanceurs d’alerte… Enfin, il proposent que les “données probantes” soient un objectif prioritaire, c’est-à-dire d’améliorer l’accès de la recherche.

Dans le domaine de l’IA, la difficulté consiste à évaluer les risques avant le déploiement. Pour améliorer la résilience, il est important d’améliorer la responsabilité et la résilience, plus que l’analyse de risque, c’est-à-dire des démarches de contrôle qui ont lieu après les déploiements. “La résilience exige à la fois de minimiser la gravité des dommages lorsqu’ils surviennent et la probabilité qu’ils surviennent.” Pour atténuer les effets de l’IA nous devons donc nous doter de politiques qui vont renforcer la démocratie, la liberté de la presse ou l’équité dans le monde du travail. C’est-à-dire d’améliorer la résilience sociétale au sens large. 

Pour élaborer des politiques technologiques efficaces, il faut ensuite renforcer les capacités techniques et institutionnelles de la recherche, des autorités et administrations. Sans personnels compétents et informés, la régulation de l’IA sera toujours difficile. Les chercheurs invitent même à “diversifier l’ensemble des régulateurs et, idéalement, à introduire la concurrence entre eux plutôt que de confier la responsabilité de l’ensemble à un seul régulateur”.

Par contre, Kapoor et Narayanan se défient fortement des politiques visant à promouvoir une non-prolifération de l’IA, c’est-à-dire à limiter le nombre d’acteurs pouvant développer des IA performantes. Les contrôles à l’exportation de matériel ou de logiciels visant à limiter la capacité des pays à construire, acquérir ou exploiter une IA performante, l’exigence de licences pour construire ou distribuer une IA performante, et l’interdiction des modèles d’IA à pondération ouverte… sont des politiques qui favorisent la concentration plus qu’elles ne réduisent les risques. “Lorsque de nombreuses applications en aval s’appuient sur le même modèle, les vulnérabilités de ce modèle peuvent être exploitées dans toutes les applications”, rappellent-ils.

Pour les deux chercheurs, nous devons “réaliser les avantages de l’IA”, c’est-à-dire accélérer l’adoption des bénéfices de l’IA et atténuer ses inconvénients. Pour cela, estiment-ils, nous devons être plus souples sur nos modalités d’intervention. Par exemple, ils estiment que pour l’instant catégoriser certains domaines de déploiement de l’IA comme à haut risque est problématique, au prétexte que dans ces secteurs (assurance, prestation sociale ou recrutement…), les technologies peuvent aller de la reconnaissance optique de caractères, relativement inoffensives, à la prise de décision automatisées dont les conséquences sont importantes. Pour eux, il faudrait seulement considérer la prise de décision automatisée dans ces secteurs comme à haut risque. 

Un autre enjeu repose sur l’essor des modèles fondamentaux qui a conduit à une distinction beaucoup plus nette entre les développeurs de modèles, les développeurs en aval et les déployeurs (parmi de nombreuses autres catégories). Une réglementation insensible à ces distinctions risque de conférer aux développeurs de modèles des responsabilités en matière d’atténuation des risques liés à des contextes de déploiement particuliers, ce qui leur serait impossible en raison de la nature polyvalente des modèles fondamentaux et de l’imprévisibilité de tous les contextes de déploiement possibles.

Enfin, lorsque la réglementation établit une distinction binaire entre les décisions entièrement automatisées et celles qui ne le sont pas, et ne reconnaît pas les degrés de surveillance, elle décourage l’adoption de nouveaux modèles de contrôle de l’IA. Or de nombreux nouveaux modèles sont proposés pour garantir une supervision humaine efficace sans impliquer un humain dans chaque décision. Il serait imprudent de définir la prise de décision automatisée de telle sorte que ces approches engendrent les mêmes contraintes de conformité qu’un système sans supervision. Pour les deux chercheurs, “opposer réglementation et diffusion est un faux compromis, tout comme opposer réglementation et innovation”, comme le disait Anu Bradford. Pour autant, soulignent les chercheurs, l’enjeu n’est pas de ne pas réguler, mais bien de garantir de la souplesse. La législation garantissant la validité juridique des signatures et enregistrement électroniques promulguée en 2000 aux Etats-Unis a joué un rôle déterminant dans la promotion du commerce électronique et sa diffusion. La législation sur les petits drones mise en place par la Federal Aviation Administration en 2016 a permis le développement du secteur par la création de pilotes certifiés. Nous devons trouver pour l’IA également des réglementations qui favorisent sa diffusion, estiment-ils. Par exemple, en facilitant “la redistribution des bénéfices de l’IA afin de les rendre plus équitables et d’indemniser les personnes qui risquent de subir les conséquences de l’automatisation. Le renforcement des filets de sécurité sociale contribuera à atténuer l’inquiétude actuelle du public face à l’IA dans de nombreux pays”. Et les chercheurs de suggérer par exemple de taxer les entreprises d’IA pour soutenir les industries culturelles et le journalisme, mis à mal par l’IA. En ce qui concerne l’adoption par les services publics de l’IA, les gouvernements doivent trouver le juste équilibre entre une adoption trop précipitée qui génère des défaillances et de la méfiance, et une adoption trop lente qui risque de produire de l’externalisation par le secteur privé.

  • ✇Dans les algorithmes
  • De l’instrumentalisation du numérique par l’action publique
    L’avant dernier numéro de la revue Réseaux est consacré à la dématérialisation. “La numérisation des administrations redéfinit le rapport à l’État et ouvre à des procès de gouvernement par instrumentation”, explique le sociologue Fabien Granjon en introduction du numéro, qui rappelle que ces développements suivent des phases en regardant celles des pays scandinaves, pionniers en la matière. Pourtant, là-bas comme ailleurs, la numérisation n’a pas aidé ou apaisé la relation administrative, au con
     

De l’instrumentalisation du numérique par l’action publique

2 juin 2025 à 01:00

L’avant dernier numéro de la revue Réseaux est consacré à la dématérialisation. “La numérisation des administrations redéfinit le rapport à l’État et ouvre à des procès de gouvernement par instrumentation”, explique le sociologue Fabien Granjon en introduction du numéro, qui rappelle que ces développements suivent des phases en regardant celles des pays scandinaves, pionniers en la matière. Pourtant, là-bas comme ailleurs, la numérisation n’a pas aidé ou apaisé la relation administrative, au contraire. Là-bas aussi, elle rend plus difficile l’accès aux droits. En fait, plus qu’un changement d’outillage, elle est une “évolution de nature politique” qui redéfinit les rapports à l’État, à la vie sociale et aux citoyens, notamment parce qu’elle s’accompagne toujours de la fermeture des guichets d’accueil physique, créant des situations d’accès inégalitaires qui restreignennt la qualité et la continuité des services publics. “La dématérialisation de l’action publique signe en cela une double délégation. D’un côté, certaines opérations sont déléguées aux dispositifs techniques ; de l’autre, on constate un déplacement complémentaire de celles-ci vers les usagers, qui portent désormais la charge et la responsabilité du bon déroulement des démarches au sein desquelles ils s’inscrivent.“ A France Travail, explique Mathilde Boeglin-Henky, les outils permettent de trier les postulants entre ceux capables de se débrouiller et les autres. L’accès aux outils numériques et leur maîtrise devient un nouveau critère d’éligibilité aux droits, générateur de non-recours : le numérique devient une “charge supplémentaire” pour les plus vulnérables. La simplification devient pour eux une “complication effective”. Pour surmonter ces difficultés, l’entraide s’impose, notamment celle des professionnels de l’accompagnement et des associations d’aide aux usagers. Mais ces nouvelles missions qui leur incombent viennent “déplacer le périmètre de leurs missions premières”, au risque de les remplacer. 

L’article de Pierre Mazet du Lab Accès sur le dispositif Conseiller numérique France Services (CnFS) montre qu’il se révèle fragile, profitant d’abord aux structures préalablement les plus engagées sur l’inclusion numérique. L’État s’est appuyé sur le plan de relance européen afin de transférer aux acteurs locaux la prise en charge d’un problème public dont les conseillers numériques doivent assumer la charge. Les moyens s’avèrent « structurellement insuffisants pour stabiliser une réponse proportionnée aux besoins ». À l’échelle nationale, les démarches en ligne se placent en tête des aides réalisées par les CnFS, montrant que « les besoins d’accompagnement sont bel et bien indexés à la numérisation des administrations »; et de constater qu’il y a là une « situation pour le moins paradoxale d’une action publique – les programmes d’inclusion numérique – qui ne parvient pas à répondre aux besoins générés par une autre action publique – les politiques de dématérialisation ». Le financement du dispositif a plus tenu d’un effet d’aubaine profitant à certains acteurs, notamment aux acteurs de la dématérialisation de la relation administrative, qu’il n’a permis de répondre à la géographie sociale des besoins. « Le dispositif a essentiellement atteint le public des personnes âgées, moins en réponse à des besoins qu’en raison du ciblage de l’offre elle-même : elle capte d’abord des publics « disponibles », pas nécessairement ceux qui en ont le plus besoin ». Enfin, la dégressivité des financements a, quant à elle, « produit un effet de sélection, qui a accentué les inégalités entre acteurs et territoires », notamment au détriment des acteurs de la médiation numérique.

La gouvernance par dispositifs numériques faciliterait l’avènement d’une administration d’orientation néolibérale priorisant les valeurs du marché, explique Granjon. L’administration « renforcerait son contrôle sur les populations, mais, paradoxalement, perdrait le contrôle sur ses principaux outils, notamment ceux d’aide à la décision quant à l’octroi de droits et de subsides ». La décision confiée aux procédures de calcul, laisse partout peu de marge de manœuvre aux agents, les transformant en simples exécutants. A Pôle Emploi, par exemple, il s’agit moins de trouver un emploi aux chômeurs que de les rendre « autonomes » avec les outils numériques. Pour Périne Brotcorne pourtant, malgré la sempiternelle affirmation d’une “approche usager”, ceux-ci sont absents des développements numériques des services publics. Rien n’est fait par exemple pour l’usager en difficulté par exemple pour qu’il puisse déléguer la prise en charge de ses tâches administratives à un tiers, comme le soulignait récemment le Défenseur des droits. Les interfaces numériques, trop complexes, fabriquent “de l’incapacité” pour certains publics, notamment les plus éloignés et les plus vulnérables. Brotcorne montre d’ailleurs très bien que “l’usager” est un concept qui permet d’avoir une “vision sommaire des publics destinataires”. Au final, les besoins s’adaptent surtout aux demandes des administrations qu’aux usagers qui ne sont pas vraiment invités à s’exprimer. L’étude souligne que près de la moitié des usagers n’arrivent pas à passer la première étape des services publics numériques que ce soit se connecter, prendre rendez-vous ou même télécharger un formulaire. Dans un autre article, signé Anne-Sylvie Pharabod et Céline Borelle, les deux chercheuses auscultent les pratiques administratives numérisées ordinaires qui montrent que la numérisation est une longue habituation, où les démarches apprises pour un service permettent d’en aborder d’autres. Les démarches administratives sont un univers de tâches dont il faut apprendre à se débrouiller, comme faire se peut, et qui en même temps sont toujours remises à zéro par leurs transformations, comme l’évolution des normes des mots de passe, des certifications d’identité, ou des documents à uploader. “La diversité des démarches, l’hétérogénéité des interfaces et l’évolution rapide des outils liée à des améliorations incrémentales (notamment en matière de sécurité) renouvellent constamment le questionnement sur ce qu’il convient de faire”.

Dans un autre article, assez complexe, Fabien Granjon explore comment l’introduction de nouveaux dispositifs numériques au sein du Service public de l’emploi a pour conséquence une reconfiguration majeure de celui-ci et provoque des changements dans les structures de relations entre acteurs. L’instrumentation numérique se voit investie de la fonction de régulation des comportements des usagers, des agents publics, mais également de bien d’autres publics, notamment tous ceux utilisant ses données et plateformes. A cette aune, France Travail est amené à devenir un « animateur d’écosystème emploi/formation/insertion » connectant divers échelons territoriaux d’intervention et une multitude d’acteurs y intervenant, comme l’expose, en partie, France Travail, via ses différentes plateformes. Granjon invite à s’intéresser à ces encastrements nouveaux et pas seulement aux guichets, ou à la relation agent-public, pour mieux saisir comment les bases de données, les API façonnent les relations avec les sous-traitants comme avec tous ceux qui interviennent depuis les procédures que France Travail met en place. 

Le numéro de Réseaux livre également un intéressant article, très critique, du Dossier médical partagé, signé Nicolas Klein et Alexandre Mathieu-Fritz, qui s’intéresse à l’histoire de la gouvernance problématique du projet, qui explique en grande partie ses écueils, et au fait que le DMP ne semble toujours pas avoir trouvé son utilité pour les professionnels de santé.

Un autre article signé Pauline Boyer explore le lancement du portail de données ouvertes de l’Etat et montre notamment que l’innovation n’est pas tant politique que de terrain. Samuel Goëta et Élise Ho-Pun-Cheung s’intéressent quant à eux à la production de standards et aux difficultés de leur intégration dans le quotidien des agents, en observant le processus de standardisation des données des lieux de médiation numérique. L’article souligne la difficulté des conseillers numériques à infléchir la standardisation et montre que cette normalisation peine à prévoir les usages de la production de données.

Dans un article de recherche qui n’est pas publié par Réseaux, mais complémentaire à son dossier, le politologue néerlandais, Pascal D. Koenig explique que le développement d’un État algorithmique modifie la relation avec les citoyens, ce qui nécessite de regarder au-delà des seules propriétés des systèmes. L’intégration des algos et de l’IA développe une relation plus impersonnelle, renforçant le contrôle et l’asymétrie de pouvoir. Pour Koenig, cela affecte directement la reconnaissance sociale des individus et le respect qu’ils peuvent attendre d’une institution. “Les systèmes d’IA, qui remplacent ou assistent l’exécution des tâches humaines, introduisent un nouveau type de représentation des agents dans les organisations gouvernementales. Ils réduisent ainsi structurellement la représentation des citoyens – en tant qu’êtres humains – au sein de l’État et augmentent les asymétries de pouvoir, en étendant unilatéralement le pouvoir informationnel de l’État”. L’utilisation de l’IA affecte également les fondements de la reconnaissance sociale dans la relation citoyen-État, liés au comportement. En tant qu’agents artificiels, les systèmes d’IA manquent de compréhension et de compassion humaines, dont l’expression dans de nombreuses interactions est un élément important pour reconnaître et traiter une personne en tant qu’individu. C’est l’absence de reconnaissance sociale qui augmente la perception de la violence administrative. “L’instauration structurelle d’une hiérarchie plus forte entre l’État et les citoyens signifie que ces derniers sont moins reconnus dans leur statut de citoyens, ce qui érode le respect que les institutions leur témoignent”. Le manque d’empathie des systèmes est l’une des principales raisons pour lesquelles les individus s’en méfient, rappelle Koenig. Or, “la numérisation et l’automatisation de l’administration réduisent les foyers d’empathie existants et contribuent à une prise en compte étroite des besoins divers des citoyens”. “Avec un gouvernement de plus en plus algorithmique, l’État devient moins capable de compréhension empathique”. Plus encore, l’automatisation de l’Etat montre aux citoyens un appareil gouvernemental où le contact humain se réduit : moins représentatif, il est donc moins disposé à prendre des décisions dans leur intérêt.

  • ✇Dans les algorithmes
  • De la maîtrise de l’automatisation d’Etat
    Le dernier numéro de la revue Multitudes (n°98, printemps 2025) publie un ensemble de contributions sur les questions algorithmiques au travers de cinq enquêtes de jeunes sociologues.  Camille Girard-Chanudet décrit la tension entre expertise algorithmique et expertise juridique dans l’essor des startups de la legal tech venues de l’extérieur des tribunaux pour les transformer.  Héloïse Eloi‑Hammer observe les différences d’implémentations des algorithmes locaux dans Parcoursup pour montre
     

De la maîtrise de l’automatisation d’Etat

28 mai 2025 à 01:00

Le dernier numéro de la revue Multitudes (n°98, printemps 2025) publie un ensemble de contributions sur les questions algorithmiques au travers de cinq enquêtes de jeunes sociologues. 

Camille Girard-Chanudet décrit la tension entre expertise algorithmique et expertise juridique dans l’essor des startups de la legal tech venues de l’extérieur des tribunaux pour les transformer. 

Héloïse Eloi‑Hammer observe les différences d’implémentations des algorithmes locaux dans Parcoursup pour montrer que les formations n’ont pas les mêmes moyens pour configurer la plateforme et que souvent, elles utilisent des procédés de sélection rudimentaires. Elle montre que, là aussi, “les algorithmes sont contraints par les contextes organisationnels et sociaux dans lesquels leurs concepteurs sont pris” et peuvent avoir des fonctionnements opposés aux valeurs des formations qui les mettent en œuvre. 

Jérémie Poiroux évoque, lui, l’utilisation de l’IA pour l’inspection des navires, une forme de contrôle technique des activités maritimes qui permet de montrer comment le calcul et l’évolution de la réglementation sont mobilisées pour réduire les effectifs des services de l’Etat tout en améliorant le ciblage des contrôles. Pour Poiroux, le système mis en place pose des questions quant à son utilisation et surtout montre que l’Etat fait des efforts pour “consacrer moins de moyens à la prévention et à l’accompagnement, afin de réduire son champ d’action au contrôle et à la punition”, ainsi qu’à éloigner les agents au profit de règles venues d’en haut. 

Soizic Pénicaud revient quant à elle sur l’histoire de la mobilisation contre les outils de ciblage de la CAF. Elle souligne que la mobilisation de différents collectifs n’est pas allé de soi et que le recueil de témoignages a permis de soulever le problème et d’incarner les difficultés auxquelles les personnes étaient confrontées. Et surtout que les collectifs ont du travailler pour “arracher la transparence”, pour produire des chiffres sur une réalité. 

Maud Barret Bertelloni conclut le dossier en se demandant en quoi les algorithmes sont des outils de gouvernement. Elle rappelle que les algorithmes n’oeuvrent pas seuls. Le déploiement des algorithmes à la CAF permet la structuration et l’intensification d’une politique rigoriste qui lui préexiste. “L’algorithme ne se substitue pas aux pratiques précédentes de contrôle. Il s’y intègre (…). Il ne l’automatise pas non plus : il le « flèche »”. Elle rappelle, à la suite du travail de Vincent Dubois dans Contrôler les assistés, que le développement des systèmes de calculs permettent à la fois de produire un contrôle réorganisé, national, au détriment de l’autonomie des agents et des caisses locales, ainsi que de légitimer la culture du contrôle et de donner une nouvelle orientation aux services publics.

Comme le murmure Loup Cellard en ouverture du dossier : “l’algorithmisation des États est le signe d’un positivisme : croyance dans la Science, confiance dans son instrumentalisme, impersonnalité de son pouvoir”.

Mardi 3 juin à 19h30 à la librairie L’atelier, 2 bis rue Jourdain, 75020 Paris, une rencontre est organisée autour des chercheurs et chercheuses qui ont participé à ce numéro. Nous y serons.

  • ✇Dans les algorithmes
  • L’IA peut-elle soutenir la démocratie ?
    La politologue Erica Chenoweth est la directrice du Non Violent Action Lab à Harvard. Elle a publié de nombreux livres pour montrer que la résistance non violente avait des effets, notamment, en français Pouvoir de la non-violence : pourquoi la résistance civile est efficace (Calmann Levy, 2021). Mais ce n’est plus le constat qu’elle dresse. Elle prépare d’ailleurs un nouveau livre, The End of People Power, qui pointe le déclin déroutant des mouvements de résistance civile au cours de la dernièr
     

L’IA peut-elle soutenir la démocratie ?

27 mai 2025 à 01:00

La politologue Erica Chenoweth est la directrice du Non Violent Action Lab à Harvard. Elle a publié de nombreux livres pour montrer que la résistance non violente avait des effets, notamment, en français Pouvoir de la non-violence : pourquoi la résistance civile est efficace (Calmann Levy, 2021). Mais ce n’est plus le constat qu’elle dresse. Elle prépare d’ailleurs un nouveau livre, The End of People Power, qui pointe le déclin déroutant des mouvements de résistance civile au cours de la dernière décennie, alors même que ces techniques sont devenues très populaires dans le monde entier. Lors d’une récente conférence sur l’IA et les libertés démocratiques organisée par le Knight First Amendment Institute de l’université de Columbia, elle se demandait si l’IA pouvait soutenir les revendications démocratiques, rapporte Tech Policy Press (vidéo). L’occasion de rendre compte à notre tour d’un point de vue décoiffant qui interroge en profondeur notre rapport à la démocratie.  

La récession démocratique est engagée

La démocratie est en déclin, explique Erica Chenoweth. Dans son rapport annuel, l’association internationale Freedom House parle même de “récession démocratique” et estime que la sauvegarde des droits démocratiques est partout en crise. 2025 pourrait être la première année, depuis longtemps, où la majorité de la population mondiale vit sous des formes de gouvernement autoritaires plutôt que démocratiques. Ce recul est dû à la fois au développement de l’autoritarisme dans les régimes autocratiques et à l’avancée de l’autocratie dans les démocraties établies, explique Chenoweth. Les avancées démocratiques des 100 dernières années ont été nombreuses et assez générales. Elles ont d’abord été le fait de mouvements non violents, populaires, où des citoyens ordinaires ont utilisé les manifestations et les grèves, bien plus que les insurrections armées, pour déployer la démocratie là où elle était empêchée. Mais ces succès sont en déclin. Les taux de réussite des mouvements populaires pacifistes comme armés, se sont effondrés, notamment au cours des dix dernières années. “Il y a quelque chose de global et de systémique qui touche toutes les formes de mobilisation de masse. Ce n’est pas seulement que les manifestations pacifiques sont inefficaces, mais que, de fait, les opposants à ces mouvements sont devenus plus efficaces pour les vaincre en général”.

Les épisodes de contestation réformistes (qui ne relèvent pas de revendications maximalistes, comme les mouvements démocratiques), comprenant les campagnes pour les salaires et le travail, les mouvements environnementaux, les mouvements pour la justice raciale, l’expansion des droits civiques, les droits des femmes… n’ont cessé de subir des revers et des défaites au cours des deux dernières décennies, et ont même diminué leur capacité à obtenir des concessions significatives, à l’image de la contestation de la réforme des retraites en France ou des mouvements écologiques, plus écrasés que jamais. Et ce alors que ces techniques de mobilisation sont plus utilisées que jamais. 

Selon la littérature, ce qui permet aux mouvements populaires de réussir repose sur une participation large et diversifiée, c’est-à-dire transversale à l’ensemble de la société, transcendant les clivages raciaux, les clivages de classe, les clivages urbains-ruraux, les clivages partisans… Et notamment quand ils suscitent le soutien de personnes très différentes les unes des autres. “Le principal défi pour les mouvements de masse qui réclament un changement pour étendre la démocratie consiste bien souvent à disloquer les piliers des soutiens autocratiques comme l’armée ou les fonctionnaires. L’enjeu, pour les mouvements démocratiques, consiste à retourner la répression à l’encontre de la population en la dénonçant pour modifier l’opinion générale ». Enfin, les mouvements qui réussissent enchaînent bien souvent les tactiques plutôt que de reposer sur une technique d’opposition unique, afin de démultiplier les formes de pression. 

La répression s’est mise à niveau

Mais l’autocratie a appris de ses erreurs. Elle a adapté en retour ses tactiques pour saper les quatre voies par lesquelles les mouvements démocratiques l’emportent. “La première consiste à s’assurer que personne ne fasse défection. La deuxième consiste à dominer l’écosystème de l’information et à remporter la guerre de l’information. La troisième consiste à recourir à la répression sélective de manière à rendre très difficile pour les mouvements d’exploiter les moments d’intense brutalité. Et la quatrième consiste à perfectionner l’art de diviser pour mieux régner”. Pour que l’armée ou la police ne fasse pas défection, les autorités ont amélioré la formation des forces de sécurité. La corruption et le financement permet de s’attacher des soutiens plus indéfectibles. Les forces de sécurité sont également plus fragmentées, afin qu’une défection de l’une d’entre elles, n’implique pas les autres. Enfin, il s’agit également de faire varier la répression pour qu’une unité de sécurité ne devienne pas une cible de mouvements populaires par rapport aux autres. Les purges et les répressions des personnes déloyales ou suspectes sont devenues plus continues et violentes. “L’ensemble de ces techniques a rendu plus difficile pour les mouvements civiques de provoquer la défection des forces de sécurité, et cette tendance s’est accentuée au fil du temps”.

La seconde adaptation clé a consisté à gagner la guerre de l’information, notamment en dominant les écosystèmes de l’information. “Inonder la zone de rumeurs, de désinformation et de propagande, dont certaines peuvent être créées et testées par différents outils d’IA, en fait partie. Il en va de même pour la coupure d’Internet aux moments opportuns, puis sa réouverture au moment opportun”.

Le troisième volet de la panoplie consiste à appliquer une répression sélective, consistant à viser des individus plutôt que les mouvements pour des crimes graves qui peuvent sembler décorrélé des manifestations, en les accusant de terrorisme ou de préparation de coup d’Etat. ”La guerre juridique est un autre outil administratif clé”.

Le quatrième volet consiste à diviser pour mieux régner. En encourageant la mobilisation loyaliste, en induisant des divisions dans les mouvements, en les infiltrant pour les radicaliser, en les poussant à des actions violentes pour générer des reflux d’opinion… 

Comment utiliser l’IA pour gagner ? 

Dans la montée de l’adaptation des techniques pour défaire leurs opposants, la technologie joue un rôle, estime Erica Chenoweth. Jusqu’à présent, les mouvements civiques ont plutôt eu tendance à s’approprier et à utiliser, souvent de manière très innovantes, les technologies à leur avantage, par exemple à l’heure de l’arrivée d’internet, qui a très tôt été utilisé pour s’organiser et se mobiliser. Or, aujourd’hui, les mouvements civiques sont bien plus prudents et sceptiques à utiliser l’IA, contrairement aux régimes autocratiques. Pourtant, l’un des principaux problèmes des mouvements civiques consiste à “cerner leur environnement opérationnel”, c’est-à-dire de savoir qui est connecté à qui, qui les soutient ou pourrait les soutenir, sous quelles conditions ? Où sont les vulnérabilités du mouvement ? Réfléchir de manière créative à ces questions et enjeux, aux tactiques à déployer pourrait pourtant être un atout majeur pour que les mouvements démocratiques réussissent. 

Les mouvements démocratiques passent bien plus de temps à sensibiliser et communiquer qu’à faire de la stratégie, rappelle la chercheuse. Et c’est là deux enjeux où les IA pourraient aider, estime-t-elle. En 2018 par exemple, lors des élections municipales russe, un algorithme a permis de contrôler les images de vidéosurveillance des bureaux de vote pour détecter des irrégularités permettant de dégager les séquences où des bulletins préremplis avaient été introduits dans les urnes. Ce qui aurait demandé un contrôle militant épuisant a pu être accompli très simplement. Autre exemple avec les applications BuyCat, BoyCott ou NoThanks, qui sont des applications de boycott de produits, permettant aux gens de participer très facilement à des actions (des applications puissantes, mais parfois peu transparentes sur leurs méthodes, expliquait Le Monde). Pour Chenoweth, les techniques qui fonctionnent doivent être mieux documentées et partagées pour qu’elles puissent servir à d’autres. Certains groupes proposent d’ailleurs déjà des formations sur l’utilisation de l’IA pour l’action militante, comme c’est le cas de Canvas, de Social Movement Technologies et du Cooperative Impact Lab.

Le Non Violent Action Lab a d’ailleurs publié un rapport sur le sujet : Comment l’IA peut-elle venir aider les mouvements démocratiques ? Pour Chenoweth, il est urgent d’évaluer si les outils d’IA facilitent ou compliquent réellement le succès des mouvements démocratiques. Encore faudrait-il que les mouvements démocratiques puissent accéder à des outils d’IA qui ne partagent pas leurs données avec des plateformes et avec les autorités. L’autre enjeu consiste à construire des corpus adaptés pour aider les mouvements à résister. Les corpus de l’IA s’appuient sur des données des 125 dernières années, alors que l’on sait déjà que ce qui fonctionnait il y a 60 ans ne fonctionne plus nécessairement.

Pourtant, estime Chenoweth, les mouvements populaires ont besoin d’outils pour démêler des processus délibératifs souvent complexes, et l’IA devrait pouvoir les y aider. “Aussi imparfait qu’ait été notre projet démocratique, nous le regretterons certainement lorsqu’il prendra fin”, conclut la politologue. En invitant les mouvements civiques à poser la question de l’utilisation de l’IA a leur profit, plutôt que de la rejetter d’emblée comme l’instrument de l’autoritarisme, elle invite à faire un pas de côté pour trouver des modalités pour refaire gagner les luttes sociales.

On lui suggérera tout de même de regarder du côté des projets que Audrey Tang mène à Taïwan avec le Collective intelligence for collective progress, comme ses « assemblées d’alignement » qui mobilise l’IA pour garantir une participation équitable et une écoute active de toutes les opinions. Comme Tang le défend dans son manifeste, Plurality, l’IA pourrait être une technologie d’extension du débat démocratique pour mieux organiser la complexité. Tang parle d’ailleurs de broad listening (« écoute élargie ») pour l’opposer au broadcasting, la diffusion de un vers tous. Une méthode mobilisée par un jeune ingénieur au poste de gouverneur de Tokyo, Takahiro Anno, qui a bénéficié d’une audience surprenante, sans néanmoins l’emporter. Son adversaire a depuis mobilisé la méthode pour lancer une consultation, Tokyo 2050.

Des pistes à observer certes, pour autant qu’on puisse mesurer vraiment leurs effets. Peuvent-elles permettent aux luttes sociales de l’emporter, comme le propose Chenoweth ? Le risque est fort de nous faire glisser vers une vie civique automatisée. Ajouter de l’IA ne signifie pas que les décisions demain seront plus justes, plus efficaces ou plus démocratiques. Au contraire. Le risque est fort que cet ajout bénéfice d’abord aux plus nantis au détriment de la diversité. L’enjeu demeure non pas d’ajouter des outils pour eux-mêmes, mais de savoir si ces outils produisent du changement et au profit de qui !

  • ✇Dans les algorithmes
  • Réutiliser, réparer, refuser, réclamer 
    Le soulèvement contre l’obsolescence programmée est bien engagé, estime Geert Lovink (blog) dans la conclusion d’un petit livre sur l’internet des choses mortes (The Internet of Dead things, édité par Benjamin Gaulon, Institute of Network Cultures, 2025, non traduit). Le petit livre, qui rassemble notamment des contributions d’artistes de l’Institut de l’internet des choses mortes, qui ont œuvré à développer un système d’exploitation pour Minitel, met en perspective l’hybridation fonctionnelle d
     

Réutiliser, réparer, refuser, réclamer 

26 mai 2025 à 01:00

Le soulèvement contre l’obsolescence programmée est bien engagé, estime Geert Lovink (blog) dans la conclusion d’un petit livre sur l’internet des choses mortes (The Internet of Dead things, édité par Benjamin Gaulon, Institute of Network Cultures, 2025, non traduit). Le petit livre, qui rassemble notamment des contributions d’artistes de l’Institut de l’internet des choses mortes, qui ont œuvré à développer un système d’exploitation pour Minitel, met en perspective l’hybridation fonctionnelle des technologies. Pour Lovink, l’avenir n’est pas seulement dans la réduction de la consommation et dans le recyclage, mais dans l’intégration à grande échelle de l’ancien dans le nouveau. Hybrider les technologies défuntes et les intégrer dans nos quotidiens est tout l’enjeu du monde à venir, dans une forme de permaculture du calcul. Arrêtons de déplorer l’appropriation du logiciel libre et ouvert par le capitalisme vautour, explique Geert Lovink. La réutilisation et la réparation nous conduisent désormais à refuser la technologie qu’on nous impose. Les mouvements alternatifs doivent désormais “refuser d’être neutralisés, écrasés et réduits au silence”, refuser de se faire réapproprier. Nous devons réclamer la tech – c’était déjà la conclusion de son précédent livre, Stuck on the platform (2022, voir notre critique) -, comme nous y invitent les hackers italiens, inspiré par le mouvement britannique des années 90, qui réclamait déjà la rue, pour reconquérir cet espace public contre la surveillance policière et la voiture.

“Reclaim the Tech » va plus loin en affirmant que « Nous sommes la technologie », explique Lovink. Cela signifie que la technologie n’est plus un phénomène passager, imposé : la technologie est en nous, nous la portons à fleur de peau ou sous la peau. Elle est intime, comme les applications menstruelles de la « femtech », décrites par Morgane Billuart dans son livre Cycles. Les ruines industrielles tiennent d’un faux romantisme, clame Lovink. Nous voulons un futur hybrid-punk, pas cypherpunk ! “La culture numérique actuelle est stagnante, elle n’est pas une échappatoire. Elle manque de direction et de destin. La volonté d’organisation est absente maintenant que même les réseaux à faible engagement ont été supplantés par les plateformes. L’esprit du temps est régressif, à l’opposé de l’accélérationnisme. Il n’y a pas d’objectif vers lequel tendre, quelle que soit la vitesse. Il n’y a pas non plus de dissolution du soi dans le virtuel. Le cloud est le nouveau ringard. Rien n’est plus ennuyeux que le virtuel pur. Rien n’est plus corporate que le centre de données. Ce que nous vivons est une succession interminable de courtes poussées d’extase orgasmique, suivies de longues périodes d’épuisement.”

Ce rythme culturel dominant a eu un effet dévastateur sur la recherche et la mise en œuvre d’alternatives durables, estime Lovink. L’optimisation prédictive a effacé l’énergie intérieure de révolte que nous portons en nous. Il ne reste que des explosions de colère, entraînant des mouvements sociaux erratiques – une dynamique alimentée par une utilisation des réseaux sociaux à courte durée d’attention. La question d’aujourd’hui est de savoir comment rendre la (post)colonialité visible dans la technologie et le design. Nous la voyons apparaître non seulement dans le cas des matières premières, mais aussi dans le contexte du « colonialisme des données ». 

Mais, s’il est essentiel d’exiger la décolonisation de tout, estime Lovink, la technologie n’abandonnera pas volontairement sa domination du Nouveau au profit de la « créolisation technologique ».

La décolonisation de la technologie n’est pas un enjeu parmi d’autres : elle touche au cœur même de la production de valeur actuelle. Prenons garde de ne pas parler au nom des autres, mais agissons ensemble, créons des cultures de « vivre ensemble hybride » qui surmontent les nouveaux cloisonnements géopolitiques et autres formes subliminales et explicites de techno-apartheid. La violence technologique actuelle va des biais algorithmiques et de l’exclusion à la destruction militaire bien réelle de terres, de villes et de vies. Les alternatives, les designs innovants, les feuilles de route et les stratégies de sortie ne manquent pas. L’exode ne sera pas télévisé. Le monde ne peut attendre la mise en œuvre des principes de prévention des données. Arrêtons définitivement les flux de données !, clame Lovink. 

La « confidentialité » des données s’étant révélée être un gouffre juridique impossible à garantir, la prochaine option sera des mécanismes intégrés, des filtres empêchant les données de quitter les appareils et les applications. Cela inclut une interdiction mondiale de la vente de données, estime-t-il. Les alternatives ne sont rien si elles ne sont pas locales. Apparaissant après la révolution, les « magasins de proximité » qui rendent les technologies aux gens ne se contenteront plus de réparer, mais nous permettront de vivre avec nos déchets, de les rendre visibles, à nouveau fonctionnels, tout comme on rend à nouveau fonctionnel le Minitel en changeant son objet, sa destination, ses modalités. 

  • ✇Dans les algorithmes
  • IA générative vs IA productive
    Encore une réflexion stimulante de Gregory Chatonsky, qui observe deux modes de relation à l’IA générative. L’un actif-productif, l’autre passif-reproductif-mémétique. « L’infrastructure générative n’a pas d’essence unifiée ni de destination prédéterminée — elle peut être orientée vers la production comme vers la consommation. Cette indétermination constitutive des technologies génératives révèle un aspect fondamental : nous nous trouvons face à un système technique dont les usages et les implic
     

IA générative vs IA productive

23 mai 2025 à 01:00

Encore une réflexion stimulante de Gregory Chatonsky, qui observe deux modes de relation à l’IA générative. L’un actif-productif, l’autre passif-reproductif-mémétique. « L’infrastructure générative n’a pas d’essence unifiée ni de destination prédéterminée — elle peut être orientée vers la production comme vers la consommation. Cette indétermination constitutive des technologies génératives révèle un aspect fondamental : nous nous trouvons face à un système technique dont les usages et les implications restent largement à définir ».

« L’enjeu n’est donc pas de privilégier artificiellement un mode sur l’autre, mais de comprendre comment ces deux rapports à la génération déterminent des trajectoires divergentes pour notre avenir technologique. En reconnaissant cette dualité fondamentale, nous pouvons commencer à élaborer une relation plus consciente et réfléchie aux technologies génératives, capable de dépasser aussi bien l’instrumentalisme naïf que le déterminisme technologique.

La génération n’est ni intrinsèquement productive ni intrinsèquement consommatrice — elle devient l’un ou l’autre selon le rapport existentiel que nous établissons avec elle. C’est dans cette indétermination constitutive que résident sa réponse à la finitude. »

  • ✇Dans les algorithmes
  • La modération n’est pas équitable
    Dans une riche et détaillée revue d’études sur la modération des plateformes du web social, Lisa Macpherson de l’association Public Knowledge démontre que cette modération n’est ni équitable ni neutre. En fait, les études ne démontrent pas qu’il y aurait une surmodération des propos conservateurs, au contraire, les contenus de droite ayant tendance à générer plus d’engagement et donc de revenus publicitaires. En fait, si certains contenus conservateurs sont plus souvent modérés, c’est parce qu’i
     

La modération n’est pas équitable

23 mai 2025 à 01:00

Dans une riche et détaillée revue d’études sur la modération des plateformes du web social, Lisa Macpherson de l’association Public Knowledge démontre que cette modération n’est ni équitable ni neutre. En fait, les études ne démontrent pas qu’il y aurait une surmodération des propos conservateurs, au contraire, les contenus de droite ayant tendance à générer plus d’engagement et donc de revenus publicitaires. En fait, si certains contenus conservateurs sont plus souvent modérés, c’est parce qu’ils enfreignent plus volontiers les règles des plateformes en colportant de la désinformation ou des propos haineux, et non du fait des biais politiques de la modération. La modération injustifiée, elle, touche surtout les communautés marginalisées (personnes racisées, minorités religieuses, femmes, LGBTQ+). Les biais de modération sont toujours déséquilibrés. Et les contenus de droite sont plus amplifiés que ceux de gauche. 

En France, rapporte Next, constatant le très faible taux de modération de contenus haineux sur Facebook, le cofondateur de l’association #jesuislà et activiste pour les droits numériques Xavier Brandao porte plainte contre Meta auprès de l’Arcom au titre du DSA. En envoyant plus de 118 signalements à Meta en quatre mois pour des discours racistes avérés, l’activiste s’est rendu compte que seulement 8 commentaires avaient été supprimés.

  • ✇Dans les algorithmes
  • ​​Droit des étrangers, modèle de la maltraitance administrative
    “Le respect des droits fondamentaux des étrangers est un marqueur essentiel du degré de protection des droits et des libertés dans un pays”, disait très clairement le Défenseur des droits en 2016. Le rapport pointait déjà très bien nombre de dérives : atteintes aux droits constantes, exclusions du droit commun, surveillances spécifiques et invasives… Près de 10 ans plus tard, contexte politique aidant, on ne sera pas étonné que ces atteintes se soient dégradées. Les réformes législatives en cont
     

​​Droit des étrangers, modèle de la maltraitance administrative

22 mai 2025 à 01:00

“Le respect des droits fondamentaux des étrangers est un marqueur essentiel du degré de protection des droits et des libertés dans un pays”, disait très clairement le Défenseur des droits en 2016. Le rapport pointait déjà très bien nombre de dérives : atteintes aux droits constantes, exclusions du droit commun, surveillances spécifiques et invasives… Près de 10 ans plus tard, contexte politique aidant, on ne sera pas étonné que ces atteintes se soient dégradées. Les réformes législatives en continue (118 textes depuis 1945, dit un bilan du Monde) n’ont pas aidé et ne cessent de dégrader non seulement le droit, mais également les services publics dédiés. Ce qu’il se passe aux Etats-Unis n’est pas une exception

Voilà longtemps que le droit des étrangers en France est plus que malmené. Voilà longtemps que les associations et autorités indépendantes qui surveillent ses évolutions le dénoncent. Des travaux des associations comme la Cimade, à ceux des autorités indépendantes, comme le Défenseur des droits ou la Cour des comptes, en passant par les travaux de chercheurs et de journalistes, les preuves ne cessent de s’accumuler montrant l’instrumentalisation mortifère de ce qui ne devrait être pourtant qu’un droit parmi d’autres. 

Couverture du livre France terre d’écueils.

Dans un petit livre simple et percutant, France, Terre d’écueils (Rue de l’échiquier, 2025, 112 pages), l’avocate Marianne Leloup-Dassonville explique, très concrètement, comment agit la maltraitance administrative à l’égard de 5,6 millions d’étrangers qui vivent en France. Marianne Leloup-Dassonville, avocate spécialisée en droit des étrangers et administratrice de l’association Droits d’urgence, nous montre depuis sa pratique les défaillances de l’administration à destination des étrangers. Une administration mise au pas, au service d’un propos politique de plus en plus déconnecté des réalités, qui concentre des pratiques mortifères et profondément dysfonctionnelles, dont nous devrions nous alarmer, car ces pratiques portent en elles des menaces pour le fonctionnement du droit et des services publics. A terme, les dysfonctionnements du droit des étrangers pourraient devenir un modèle de maltraitance à appliquer à tous les autres services publics. 

Le dysfonctionnement des services publics à destination des étrangers est caractérisé par plusieurs maux, détaille Marianne Leloup-Dassonville. Ce sont d’abord des services publics indisponibles, ce sont partout des droits à minima et qui se réduisent, ce sont partout des procédures à rallonge, où aucune instruction n’est disponible dans un délai “normal”. Ce sont enfin une suspicion et une persécution systématique. L’ensemble produit un empêchement organisé qui a pour but de diminuer l’accès au droit, d’entraver la régularisation. Ce que Marianne Leloup-Dassonville dessine dans son livre, c’est que nous ne sommes pas seulement confrontés à des pratiques problématiques, mais des pratiques systémiques qui finissent par faire modèle. 

Indisponibilité et lenteur systémique

France, Terre d’écueils décrit d’abord des chaînes de dysfonctionnements administratifs. Par des exemples simples et accessibles, l’avocate donne de l’épaisseur à l’absurdité qui nous saisit face à la déshumanisation des étrangers dans les parcours d’accès aux droits. Nul n’accepterait pour quiconque ce que ces services font subir aux publics auxquels ils s’adressent. L’une des pratiques les plus courantes dans ce domaine, c’est l’indisponibilité d’un service : service téléphonique qui ne répond jamais, site web de prise de rendez-vous sans proposition de rendez-vous, dépôts sans récépissés, dossier auquel il n’est pas possible d’accéder ou de mettre à jour… Autant de pratiques qui sont là pour faire patienter c’est-à-dire pour décourager… et qui nécessitent de connaître les mesures de contournements qui ne sont pas dans la procédure officielle donc, comme de documenter l’indisponibilité d’un site par la prise de capture d’écran répétée, permettant de faire un recours administratif pour obliger le service à proposer un rendez-vous à un requérant. 

Toutes les procédures que l’avocate décrit sont interminables. Toutes oeuvrent à décourager comme si ce découragement était le seul moyen pour désengorger des services partout incapables de répondre à l’afflux des demandes. Si tout est kafkaïen dans ce qui est décrit ici, on est surtout marqué par la terrible lenteur des procédures, rallongées de recours eux-mêmes interminables. Par exemple, l’Ofpra (Office Français de Protection des Réfugiés et Apatrides) statue en moyenne sur une demande d’asile en 4 mois, mais souvent par un refus. Pour une procédure de recours, la CNDA (Cour nationale du droit d’asile) met en moyenne 12 à 24 mois à réétudier un dossier. Quand un ressortissant étranger obtient le statut de réfugié, il devrait obtenir un titre de séjour dans les 3 mois. En réalité, les cartes de séjour et les actes d’état civils mettent de longs mois à être produits (douze mois en moyenne) ce qui prive les étrangers de nombreux droits sociaux, comme de pouvoir s’inscrire à France Travail ou de bénéficier des allocations familiales, du droit du travail. Même constat à l’AES (admission exceptionnelle au séjour). Depuis 2023, la demande d’AES se fait en ligne et consiste à enregistrer quelques documents pour amorcer la procédure. Mais les préfectures ne font de retour que 6 mois après le dépôt de la demande. Un rendez-vous n’est proposé bien souvent qu’encore une année plus tard, soit en moyenne 18 mois entre la demande initiale et le premier rendez-vous, et ce uniquement pour déposer un dossier. L’instruction qui suit, elle, prend entre 18 et 24 mois. Il faut donc compter à minima 3 à 4 années pour une demande de régularisation ! Sans compter qu’en cas de refus (ce qui est plutôt majoritairement le cas), il faut ajouter la durée de la procédure de recours au tribunal administratif. Même chose en cas de refus de visa injustifié, qui nécessite plus de 2 ans de procédure pour obtenir concrètement le papier permettant de revenir en France. 

C’est un peu comme s’il fallait patienter des années pour obtenir un rendez-vous chez un dermatologue pour vérifier un grain de beauté inquiétant ou plusieurs mois pour obtenir des papiers d’identité. Dans le domaine du droit des étrangers, la dégradation des services publics semble seulement plus prononcée qu’ailleurs.

Droits minimaux et décisions discrétionnaires

Le dysfonctionnement que décrit l’avocate n’est pas qu’administratif, il est également juridique. Dans le domaine de la demande d’asile par exemple, il n’existe pas de procédure d’appel, qui est pourtant une garantie simple et forte d’une justice équitable. Les demandeurs ayant été refusés peuvent demander un réexamen qui a souvent lieu dans le cadre de procédures « accélérées ». Mais cette accélération là ne consiste pas à aller plus vite, elle signifie seulement que la décision sera prise par un juge unique, et donc, potentiellement, moins équitable.

L’admission exceptionnelle au séjour (AES), autre possibilité pour obtenir une naturalisation, nécessite elle de faire une demande à la préfecture de son domicile. Même si cette admission est bornée de conditions concrètes très lourdes (promesse d’embauche, présence sur le territoire depuis 3 ans, bulletins de salaires sur au moins 2 ans…), le préfet exerce ici une compétence discrétionnaire, puisqu’il évalue seul les demandes. « En matière de naturalisation, l’administration a un très large pouvoir d’appréciation », souligne l’avocate. Même constat quant aux demandes de naturalisation, tout aussi compliquées que les autres demandes. L’entretien de naturalisation par exemple, qu’avait étudié le sociologue Abdellali Hajjat dans son livre, Les frontières de l’identité nationale, a tout d’une sinécure. Il consiste en un entretien discrétionnaire, sans procès verbal ni enregistrement, sans programme pour se préparer à y répondre. Ce n’est pas qu’un test de culture général d’ailleurs auquel nombre de Français auraient du mal à répondre, puisque de nombreuses questions ne relèvent pas d’un QCM, mais sont appréciatives. Il n’y a même pas de score minimal à obtenir, comme ce peut-être le cas dans d’autres pays, comme le Royaume-Uni. Vous pouvez répondre juste aux questions… et être refusé.

Suspicion, persécution et empêchement systématiques

Le fait de chercher à empêcher les demandeurs d’asile d’accéder à un travail (en janvier 2024, une loi anti-immigration, une de plus, est même venue interdire à ceux qui ne disposent pas de titre de séjour d’obtenir un statut d’entrepreneur individuel) créé en fait une surcouche de dysfonctionnements et de coûts qui sont non seulement contre-productifs, mais surtout rendent la vie impossible aux étrangers présents sur le territoire. Et ce alors que 90% des personnes en situation irrégulière travaillent quand même, comme le souligne France Stratégies dans son rapport sur l’impact de l’immigration sur le marché du travail. En fait, en les empêchant de travailler, nous produisons une surcouche de violences : pour travailler, les sans-papiers produisent les titres de séjour d’un tiers, à qui seront adressés les bulletins de salaires. Tiers qui prélèvent une commission de 30 à 50% sur ces revenus, comme le montrait l’excellent film de Boris Lokjine, L’histoire de Souleymane.

Parce que le lien de paternité ou de maternité créé un droit pour l’enfant et pour le compagnon ou la compagne, la suspicion de mariage blanc est devenue systématique quand elle était avant un signe d’intégration, comme le rappelait le rapport du mouvement des amoureux aux ban publics. Enquêtes intrusives, abusives, surveillance policière, maires qui s’opposent aux unions (sans en avoir le droit)… Alors qu’on ne comptait que 345 mariages annulés en 2009 (soit moins de 0,5% des unions mixtes) et une trentaine de condamnations pénales par an, 100% des couples mixtes subissent des discriminations administratives. Le soupçon mariage blanc permet surtout de masquer les défaillances et le racisme de l’administration française. Ici comme ailleurs, le soupçon de fraude permet de faire croire à tous que ce sont les individus qui sont coupables des lacunes de l’administration française. 

Une politique du non accueil perdu dans ses amalgames

France, terre d’écueils, nous montre le parcours du combattant que représente partout la demande à acquérir la nationalité. L’ouvrage nous montre ce que la France impose à ceux qui souhaitent y vivre, en notre nom. Une politique de non accueil qui a de quoi faire honte. 

Ces dérives juridiques et administratives sont nées de l’amalgame sans cesse renouvelé à nous faire confondre immigration et délinquance. C’est ce que symbolise aujourd’hui le délires sur les OQTF. L’OQTF est une décision administrative prise par les préfectures à l’encontre d’un étranger en situation irrégulière, lui imposant un départ avec ou sans délai. Or, contrairement à ce que l’on ne cesse de nous faire croire, les OQTF ne concernent pas uniquement des individus dangereux, bien au contraire. La grande majorité des personnes frappées par des OQTF ne sont coupables d’aucun délit, rappelle Marianne Leloup-Dassonville. Elles sont d’abord un objectif chiffré. Les préfectures les démultiplent alors qu’elles sont très souvent annulées par les tribunaux. La France en émet plus de 100 000 par an et en exécute moins de 7000. Elle émet 30% des OQTF de toute l’Europe. Les services préfectoraux dédiés sont surchargés et commettent des “erreurs de droit” comme le dit très pudiquement la Cour des comptes dans son rapport sur la lutte contre l’immigration irrégulière. Le contentieux des étrangers représente 41% des affaires des tribunaux administratifs en 2023 (contre 30% en 2016) et 55% des affaires des cours d’appel, comme le rappelait Mediapart. La plupart des OQTF concernent des ressortissants étrangers qui n’ont jamais commis d’infraction (ou des infractions mineures et anciennes). Nombre d’entre elles sont annulées, nombre d’autres ne peuvent pas être exécutées. Peut-on sortir de cette spirale inutile, se demande l’avocate ?

Nous misons désormais bien plus sur une fausse sécurité que sur la sûreté. La sûreté, rappelle l’avocate, c’est la garantie que les libertés individuelles soient respectées, contre une arrestation, un emprisonnement ou une condamnation arbitraire. La sûreté nous garantit une administration équitable et juste. C’est elle que nous perdons en nous enfermant dans un délire sécuritaire qui a perdu contact avec la réalité. Les témoignages que livre Marianne Leloup-Dassonville montrent des personnes plutôt privilégiées, traumatisées par leurs rapports à l’administration française. Les étrangers sont partout assimilés à des délinquants comme les bénéficiaires de droits sociaux sont assimilés à des fraudeurs.

A lire le court ouvrage de Marianne Leloup-Dassonville, on se dit que nous devrions nous doter d’un observatoire de la maltraitance administrative pour éviter qu’elle ne progresse et qu’elle ne contamine toutes les autres administrations. Nous devons réaffirmer que l’administration ne peut se substituer nulle part à la justice, car c’est assurément par ce glissement là qu’on entraîne tous les autres. 

Hubert Guillaud

PS : Au Royaume-Uni, le ministère de l’intérieur souhaite introduire deux outils d’IA dans les processus de décisions de demande d’asile d’après une étude pilote : une pour résumer les informations des dossiers, l’autre pour trouver des informations sur les pays d’origine des demandeurs d’asile. Une façon de renforcer là-bas, la boîte noire des décisions, au prétexte d’améliorer « la vitesse » de décision plutôt que leur qualité, comme le souligne une étude critique. Comme quoi, on peut toujours imaginer pire pour encore dégrader ce qui l’est déjà considérablement.

MAJ du 23/05/2025 : L’extension des prérogatives de l’administration au détriment du droit, permettant de contourner la justice sur l’application de mesures restreignant les libertés individuelles, à un nom, nous apprend Blast, l’« administrativisation », et permet des sanctions sans garanties de justice et d’appliquer des sanctions sans preuves.

  • ✇Dans les algorithmes
  • Regouverner la monétisation ?
    Pour Tech Policy Press, Victoire Rio de What to Fix estime qu’il faut réguler les programmes de monétisation de contenus que proposent désormais toutes les plateformes. Facebook a plus de 3,8 millions d’éditeurs « partenaires », comme il les appelle. Et plus d’un million ont été créés en 2025 seulement. Ces modèles de redistribution des revenus sont souvent très simples : plus le nombre de vues et d’engagements générés sont élevés, plus la plateforme verse d’argent. Or, la communauté de la régul
     

Regouverner la monétisation ?

21 mai 2025 à 01:00

Pour Tech Policy Press, Victoire Rio de What to Fix estime qu’il faut réguler les programmes de monétisation de contenus que proposent désormais toutes les plateformes. Facebook a plus de 3,8 millions d’éditeurs « partenaires », comme il les appelle. Et plus d’un million ont été créés en 2025 seulement. Ces modèles de redistribution des revenus sont souvent très simples : plus le nombre de vues et d’engagements générés sont élevés, plus la plateforme verse d’argent. Or, la communauté de la régulation s’est beaucoup intéressée à la modération des contenus, aux recommandations publicitaires, aux problèmes que posent l’amplification algorithmique, mais elle s’est peu penchée sur la « gouvernance de leur monétisation », c’est-à-dire à la façon dont les plateformes redistribuent l’argent et à la manière dont leurs décisions de monétisation « façonnent in fine l’environnement informationnel ».

Dans un rapport consacré au sujet, What to Fix estime que « les plateformes de médias sociaux ont redistribué plus de 20 milliards de dollars de leurs revenus à plus de 6 millions de comptes l’année dernière ». « Et pourtant, nous ignorons qui les plateformes ont payé, combien elles ont payé, ni pour quel contenu ! », rappelle Victoire Rio. L’enjeu n’est pas que cette redistribution n’ait pas lieu, bien au contraire, précise-t-elle : « Ce qui m’inquiète, c’est de laisser les décisions de redistribution des revenus aux plateformes de médias sociaux, sans transparence ni contrôle ». Et notamment, le fait que ces revenus bénéficient de plus en plus à des acteurs de la désinformation ou encore le fait que ces financements puissent être coupés unilatéralement pour nombre d’acteurs légitimes, avec des recours limités et sans compensation pour les pertes de revenus associées. Or, ni le DSA européen, ni le Online Safety Act britannique, n’abordent clairement les systèmes de monétisation des plateformes comme un vecteur de risque. Là encore, la transparence des modalités de reversement est en cause, accuse Victoire Rio, qui souhaite que nous considérions les modalités de monétisation comme un enjeu spécifique, c’est-à-dire un enjeu de régulation.

Cette brève est parue originellement dans la lettre du Conseil national du numérique du 7 mai 2025.

❌
❌