Discord is facing backlash after announcing that all users will soon be required to verify ages to access adult content by sharing video selfies or uploading government IDs.

According to Discord, it's relying on AI technology that verifies age on the user's device, either by evaluating a user's facial structure or by comparing a selfie to a government ID. Although government IDs will be checked off-device, the selfie data will never leave the user's device, Discord emphasized. Both forms of data will be promptly deleted after the user's age is estimated.

In a blog, Discord confirmed that "a phased global rollout" would begin in "early March," at which point all users globally would be defaulted to "teen-appropriate" experiences.

To unblur sensitive media or access age-restricted channels, the majority of users will likely have to undergo Discord's age estimation process. Most users will only need to verify their ages once, Discord said, but some users "may be asked to use multiple methods, if more information is needed to assign an age group," the blog said.

On social media, alarmed Discord users protested the move, doubting whether Discord could be trusted with their most sensitive information after Discord age verification data was recently breached. In October, hackers stole government IDs of 70,000 Discord users from a third-party service that Discord previously trusted to verify ages in the United Kingdom and Australia.

At that time, Discord told users that the hackers were hoping to use the stolen data to "extort a financial ransom from Discord." In October, Ars Senior Security Editor Dan Goodin joined others warning that "the best advice for people who have submitted IDs to Discord or any other service is to assume they have been or soon will be stolen by hackers and put up for sale or used in extortion scams."

For bad actors, Discord will likely only become a bigger target as more sensitive information is collected worldwide, users now fear.

It's no surprise then that hundreds of Discord users on Reddit slammed the decision to expand age verification globally shortly after The Verge broke the news. On a PC gaming subreddit discussing alternative apps for gamers, one user wrote, "Hell, Discord has already had one ID breach, why the fuck would anyone verify on it after that?"

"This is how Discord dies," another user declared. "Seriously, uploading any kind of government ID to a 3rd party company is just asking for identity theft on a global scale."

Many users seem just as sketched out about sharing face scans. On the Discord app subreddit, some users vowed to never submit selfies or IDs, fearing that breaches may be inevitable and suspecting Discord of downplaying privacy risks while allowing data harvesting.

Who can access Discord age-check data?

Discord's system is supposed to make sure that only users have access to their age-check data, which Discord said would never leave their phones.

The company is hoping to convince users that it has tightened security after the breach by partnering with k-ID, an increasingly popular age-check service provider that's also used by social platforms from Meta and Snap.

However, self-described Discord users on Reddit aren’t so sure, with some going the extra step of picking apart k-ID's privacy policy to understand exactly how age is verified without data ever leaving the device.

"The wording is pretty unclear and inconsistent even if you dig down to the k-ID privacy policy," one Redditor speculated. "Seems that ID scans are uploaded to k-ID servers, they delete them, but they also mention using 'trusted 3rd parties' for verification, who may or may not delete it." That user seemingly gave up on finding reassurances in either company's privacy policies, noting that "everywhere along the chain it reads like 'we don't collect your data, we forward it to someone else... .'"

To better understand user concerns, Ars reviewed the privacy policies, noting that k-ID said its "facial age estimation" tool is provided by a Swiss company called Privately.

"We don’t actually see any faces that are processed via this solution," k-ID's policy said.

That part does seem vague, since Privately isn't explicitly included in the "we" in that statement. Similarly, further down, the policy more clearly states that "neither k-ID nor its service providers collect any biometric information from users when they interact with the solution. k-ID only receives and stores the outcome of the age check process." In that section, "service providers" seems to refer to partners like Discord, which integrate k-ID's age checks, rather than third parties like Privately that actually conduct the age check.

Asked for comment, a k-ID spokesperson told Ars that "the Facial Age Estimation technology runs entirely on the user's device in real time when they are performing the verification. That means there is no video or image transmitted, and the estimation happens locally. The only data to leave the device is a pass/fail of the age threshold which is what Discord receives (and some performance metrics that contain no personal data)."

K-ID's spokesperson told Ars that no third parties store personal data shared during age checks.

"k-ID, does not receive personal data from Discord when performing age-assurance," k-ID's spokesperson said. "This is an intentional design choice grounded in data protection and data minimisation principles. There is no storage of personal data by k-ID or any third parties, regardless of the age assurance method used."

Turning to Privately's website, that offers a little more information on how on-device age estimation works, while providing likely more reassurances that data won't leave devices.

Privately's services were designed to minimize data collection and prioritize anonymity to comply with the European Union's General Data Protection Regulation, Privately noted. "No user biometric or personal data is captured or transmitted," Privately's website said, while bragging that "our secret sauce is our ability to run very performant models on the user device or user browser to implement a privacy-centric solution."

The company’s privacy policy offers slightly more detail, noting that the company avoids relying on the cloud while running AI models on local devices.

"Our technology is built using on-device edge-AI that facilitates data minimization so as to maximise user privacy and data protection," the privacy policy said. "The machine learning based technology that we use (for age estimation and safeguarding) processes user’s data on their own devices, thereby avoiding the need for us or for our partners to export user’s personal data onto any form of cloud services."

Additionally, the policy said, "our technology solutions are built to operate mostly on user devices and to avoid sending any of the user's personal data to any form of cloud service. For this we use specially adapted machine learning models that can be either deployed or downloaded on the user’s device. This avoids the need to transmit and retain user data outside the user device in order to provide the service."

Finally, Privately explained that it also employs a "double blind" implementation to avoid knowing the origin of age estimation requests. That supposedly ensures that Privately only knows the result of age checks and cannot connect the result to a user on a specific platform.

Asked for comment, Discord's spokesperson said that "Discord and our age assurance vendor partners do not permanently store personal identity documents or users’ video selfies. Identity documents, including selfies, are deleted once a user’s age group is confirmed, and the selfie video used for facial age estimation never leaves their device."

"We're also exploring other vendors and will be transparent with users if the data practices for vendors differ," Discord's spokesperson said. "We'll continue to put user privacy first as we consider introducing any additional methods in the future. We also frequently audit our third-party systems to ensure they meet our security and privacy standards."

Discord expects to lose users

Some Discord users may never be asked to verify their ages, even if they try to access age-restricted content. Savannah Badalich, Discord’s global head of product policy, told The Verge that Discord "is also rolling out an age inference model that analyzes metadata, like the types of games a user plays, their activity on Discord, and behavioral signals like signs of working hours or the amount of time they spend on Discord."

"If we have a high confidence that they are an adult, they will not have to go through the other age verification flows,” Badalich said.

Badalich confirmed that Discord is bracing for some users to leave Discord over the update but suggested that "we’ll find other ways to bring users back."

On Reddit, Discord users complained that age verification is easy to bypass, forcing adults to share sensitive information without keeping kids away from harmful content. In Australia, where Discord's policy first rolled out, some kids claimed that Discord never even tried to estimate their ages, while others found it easy to trick k-ID by using AI videos or altering their appearances to look older. A teen girl relied on fake eyelashes to do the trick, while one 13-year-old boy was estimated to be over 30 years old after scrunching his face to seem more wrinkled.

Badalich told The Verge that Discord doesn't expect the tools to work perfectly but acts quickly to block workarounds, like teens using Death Stranding's photo mode to skirt age gates. However, questions remain about the accuracy of Discord's age estimation model in assessing minors' ages, in particular.

It may be noteworthy that Privately only claims that its technology is "proven to be accurate to within 1.3 years, for 18-20-year-old faces, regardless of a customer’s gender or ethnicity." But experts told Ars last year that flawed age-verification technology still frequently struggles to distinguish minors from adults, especially when differentiating between a 17- and 18-year-old, for example.

Perhaps notably, Discord's prior scandal occurred after hackers stole government IDs that users shared as part of the appeal process in order to fix an incorrect age estimation. Appeals could remain the most vulnerable part of this process, The Verge's report indicated. Badalich confirmed that a third-party vendor would be reviewing appeals, with the only reassurance for users seemingly that IDs shared during appeals “are deleted quickly—in most cases, immediately after age confirmation.”

On Reddit, Discord fans awaiting big changes remain upset. A disgruntled Discord user suggested that "corporations like Facebook and Discord, will implement easily passable, cheapest possible, bare minimum under the law verification, to cover their ass from a lawsuit," while forcing users to trust that their age-check data is secure.

Another user joked that she'd be more willing to trust that selfies never leave a user's device if Discord were "willing to pay millions to every user" whose "scan does leave a device."

This story was updated on February 9 to add comments from Discord and k-ID, and to clarify that government IDs are checked off-device.

Read full article

Comments

Sure, most Americans are glued to their TVs for the today's Super Bowl and/or the Winter Olympics. But for the non-sports minded, Amazon MGM Studios has released one last trailer for its forthcoming space odyssey Project Hail Mary, based on Andy Weir’s (The Martian) bestselling 2021 novel about an amnesiac biologist-turned-schoolteacher in space.

As previously reported, Amazon MGM Studios acquired the rights for Weir’s novel before it was even published and brought on Drew Goddard to write the screenplay. (Goddard also wrote the adapted screenplay for The Martian, so he’s an excellent choice.) The studio tapped Phil Lord and Christopher Miller (Cloudy with a Chance of Meatballs, The LEGO Movie) to direct and signed on Ryan Gosling to star. Per the official premise:

Science teacher Ryland Grace (Ryan Gosling) wakes up on a spaceship light years from home with no recollection of who he is or how he got there. As his memory returns, he begins to uncover his mission: solve the riddle of the mysterious substance causing the sun to die out. He must call on his scientific knowledge and unorthodox ideas to save everything on Earth from extinction… but an unexpected friendship means he may not have to do it alone.

In addition to Gosling, the cast includes Sandra Huller as head of the Hail Mary project and Ryland’s superior; Milana Vayntrub as project astronaut Olesya Ilyukhina; Ken Leung as project astronaut Yao Li-Jie; Liz Kingsman as Shapiro; Orion Lee as Xi; and James Ortiz as a new life form Ryland names Rocky.

To say there's a lot of interest in this movie might be an understatement. The first trailer was released in June and racked up a whopping 400 million views worldwide in its first week. The footage—which included Ryland discovering an alien ship inhabited by the aforementioned Rocky—gave every indication of following Weir’s novel pretty closely. That’s very good news for Weir fans, which includes several of us here at Ars.

That earlier trailer mostly gave us a lot of backstory about how Ryland ended up reluctantly agreeing to the mission, with just a few glimpses of Rocky. But Rocky is front and center for this latest one. The footage focuses on how Ryland and Rocky learn to communicate and gradually bond over their shared fates, with strains of Prince's "I Would Die For You" echoing in the background. It starts with Rocky imitating Ryland's body motions, including a goofy hero pose. Eventually Ryland figures out how to synthesize and voice for Rocky so they can better coordinate their strategies. Will they succeed? Or is this a one-way trip for one or both of them?

Project Hail Mary hits theaters on March 20, 2026.

Read full article

Comments

Last week, filmmaker Darren Aronofsky's AI studio Primordial Soup and Time magazine released the first two episodes of On This Day... 1776. The year-long series of short-form videos features short vignettes describing what happened on that day of the American Revolution 250 years ago, but it does so using “a variety of AI tools” to produce photorealistic scenes containing avatars of historical figures like George Washington, Thomas Paine, and Benjamin Franklin.

In announcing the series, Time Studios President Ben Bitonti said the project provides "a glimpse at what thoughtful, creative, artist-led use of AI can look like—not replacing craft but expanding what’s possible and allowing storytellers to go places they simply couldn’t before."

Outside critics were decidedly less excited about the effort. The AV Club took the introductory episodes to task for "repetitive camera movements [and] waxen characters" that make for "an ugly look at American history." CNET said that this "AI slop is ruining American history," calling the videos a "hellish broth of machine-driven AI slop and bad human choices." The Guardian lamented that the "once-lauded director of Black Swan and The Wrestler has drowned himself in AI slop," calling the series "embarrassing," "terrible," and "ugly as sin." I could go on.

But this kind of initial reaction apparently hasn't deterred Primordial Soup from its still-evolving efforts. A source close to the production, who requested anonymity to speak frankly about details of the series' creation, told Ars that the quality of new episodes would improve as the team's AI tools are refined throughout the year and as the team learns to better use them.

"We're going into this fully assuming that we have a lot to learn, that this process is gonna evolve, the tools we're using are gonna evolve," the source said. "We're gonna make mistakes. We're gonna learn a lot... we're going to get better at it, [and] the technology will change. We'll see how audiences are reacting to certain things, what works, what doesn't work. It's a huge experiment, really."

Not all AI

It's important to note that On This Day... 1776 is not fully crafted by AI. The script, for instance, was written by a team of writers overseen by Aronofsky's longtime writing partners Ari Handel and Lucas Sussman, as noted by The Hollywood Reporter. That makes criticisms like the Guardian's of "ChatGPT-sounding sloganeering" in the first episodes both somewhat misplaced and hilariously harsh.

Our production source says the project was always conceived as a human-written effort and that the team behind it had long been planning and researching how to tell this kind of story. "I don't think [they] even needed that kind of help or wanted that kind of [AI-powered writing] help," they said. "We've all experimented with [AI-powered] writing and the chatbots out there, and you know what kind of quality you get out of that."

What you see here is not a real human actor, but his lines were written and voiced by humans. Credit: Primordial Soup

The producers also go out of their way to note that all the dialogue in the series is recorded directly by Screen Actors Guild voice actors, not by AI facsimiles. While recently negotiated union rules might have something to do with that, our production source also said the AI-generated voices the team used for temp tracks were noticeably artificial and not ready for a professional production.

Humans are also directly responsible for the music, editing, sound mixing, visual effects, and color correction for the project, according to our source. The only place the "AI-powered tools" come into play is in the video itself, which is crafted with what the announcement calls a "combination of traditional filmmaking tools and emerging AI capabilities."

In practice, our source says, that means humans create storyboards, find visual references for locations and characters, and set up how they want shots to look. That information, along with the script, gets fed into an AI video generator that creates individual shots one at a time, to be stitched together and cleaned up by humans in traditional post-production.

That process takes the AI-generated cinema conversation one step beyond Ancestra, a short film Primordial Soup released last summer in association with Google DeepMind (which is not involved with the new project). There, AI tools were used to augment "live-action scenes with sequences generated by Veo."

"Weeks" of prompting and re-prompting

In theory, having an AI model generate a scene in minutes might save a lot of time compared to traditional filmmaking—scouting locations, hiring actors, setting up cameras and sets, and the like. But our production source said the highly iterative process of generating and perfecting shots for On This Day... 1776 still takes "weeks" for each minutes-long video and that "more often than not, we're pushing deadlines."

Even though the AI model is essentially animating photorealistic avatars, the source said the process is "more like live action filmmaking" because of the lack of fine-grained control over what the video model will generate. "You don't know if you're gonna get what you want on the first take or the 12th take or the 40th take," the source said.

While some shots take less time to get right than others, our source said the AI model rarely produces a perfect, screen-ready shot on the first try. And while some small issues in an AI-generated shot can be papered over in post-production with visual effects or careful editing, most of the time, the team has to go back and tell the model to generate a completely new video with small changes.

"It still takes a lot of work, and it's not necessarily because it's wrong, per se, so much as trying to get the right control because you [might] want the light to land on the face in the right way to try to tell the story," the source said. "We're still, we're still striving for the same amount of control that we always have [with live-action production] to really maximize the story and the emotion."

Quick shots and smaller budgets

Though video models have advanced since the days of the nightmarish clip of Will Smith eating spaghetti, hallucinations and nonsensical images are "still a problem" in producing On This Day... 1776, according to our source. That's one of the reasons the company decided to use a series of short-form videos rather than a full-length movie telling the same essential story.

"It's one thing to stay consistent within three minutes. It's a lot harder and it takes a lot more work to stay consistent within two hours," the source said. "I don't know what the upper limit is now [but] the longer you get, the more things start to fall off."

We've come a long way from the circa-2023 videos of Will Smith eating spaghetti. Credit: chaindrop / Reddit

Keeping individual shots short also allows for more control and fewer "reshoots" for an AI-animated production like this. "When you think about it, if you're trying to create a 20-second clip, you have all these things that are happening, and if one of those things goes wrong in 20 seconds, you have to start over," our source said. "And the chance of something going wrong in 20 seconds is pretty high. The chance of something going wrong in eight seconds is a lot lower."

While our production source couldn't give specifics on how much the team was spending to generate so much AI-modeled video, they did suggest that the process was still a good deal cheaper than filming a historical docudrama like this on location.

"I mean, we could never achieve what we're doing here for this amount of money, which I think is pretty clear when you watch this," they said. In future episodes, the source promised, "you'll see where there's things that cameras just can't even do" as a way to "make the most of that medium."

"Let's see what we can do"

If you've been paying attention to how fast things have been moving with AI-generated video, you might think that AI models will soon be able to produce Hollywood-quality cinema with nothing but a simple prompt. But our source said that working on On This Day... 1776 highlights just how important it is for humans to still be in the loop on something like this.

"Personally, I don't think we're ever gonna get there [replacing human editors]," he said. "We actually desperately need an editor. We need another set of eyes who can look at the cut and say, 'If we get out of this shot a little early, then we can create a little bit of urgency. If we linger on this thing a little longer...' You still really need that."

AI Ben Franklin and AI Thomas Paine toast to the war propaganda effort. Credit: Primordial Soup

That could be good news for human editors. But On This Day... 1776 also suggests a world where on-screen (or even motion-captured) human actors are fully replaced by AI-generated avatars. When I asked our source why the producers felt that AI was ready to take over that specifically human part of the film equation, though, the response surprised me.

"I don't know that we do know that, honestly," they said. "I think we know that the technology is there to try. And I think as storytellers we're really interested in using... all the different tools that we can to try to get our story across and to try to make audiences feel something."

"It's not often that we have huge new tools like this," the source continued. "I mean, it's never happened in my lifetime. But when you do [get these new tools], you want to start playing with them... We have to try things in order to know if it works, if it doesn't work."

"So, you know, we have the tools now. Let's see what we can do."

Read full article

Comments

As the 2026 Olympic Winter Games begin today, news articles are swelling with juicy claims that male ski jumpers have injected their penises with fillers to gain a flight advantage.

As the rumor goes, having a bigger bulge on a required 3D body scan taken in the pre-season could earn jumpers extra centimeters of material in their jumpsuits—and a suit's larger nether regions provide more surface area to glide to the gold. Even a small increase can make a satisfying difference in this sport. A 2025 simulation-based study published in the journal Frontiers in Sports and Active Living suggested that every 2 cm of extra fabric in a ski jumpsuit could increase drag by about 4 percent and increase lift by about 5 percent. On a jump, that extra 2 cm of fabric amounts to an extra 5.8 meters, the simulations found.

Elite ski jumpers are aware of the advantage and have already crotch-rocketed to scandal with related schemes. Last year, two Norwegian Olympic medalists, Marius Lindvik and Johann Andre Forfang, and three of their team officials were charged with cheating after an anonymous video showed the head coach and suit technician illegally restitching the crotch area of the two jumpers' suits to make them larger. The jumpers received a three-month suspension, while the head coach, an assistant coach, and the technician faced a harsher 18-month ban.

Injections are alleged to be a new, more drastic strategy. Rumors that jumpers were internally padding their peckers first came to light in January, when German newspaper Bild reported that there were "whispers" of jumpers using hyaluronic acid or possibly paraffin injections.

Bild quoted Dr. Kamran Karim, a specialist at Maria-Hilf Hospital in Krefeld, Germany, as saying (translated): "There is the possibility of obtaining a temporary, optical thickening of the penis with the injection of paraffin or hyaluronic acid. ... Such an injection, however, is not medically indicated and involves risks."

On Thursday, the injection claims sprang up again at a press conference in which journalists asked officials of the World Anti-Doping Agency about the claims. The agency's director general said they were not aware of any claims and that non-doping means of enhancing performance are not in their purview. But according to the BBC, WADA President Witold Banka, who is from Poland, was "clearly entertained" by the questions, responding: "Ski jumping is very popular in Poland, so I promise you I'm going to look at it."

Bruno Sassi, the communications director for FIS, the international ski and snowboard federation, seemed less amused, telling the BBC, "There has never been any indication, let alone evidence, that any competitor has ever made use of a hyaluronic acid injection to attempt to gain a competitive advantage."

But what if they did? Here's what we know about hyaluronic acid and paraffin for penis augmentation.

Hyaluronic acid

While some news outlets have played up the "acid" part of its name, hyaluronic acid is not some nefarious flesh-melting hazard. It's a common filler used for various clinical purposes.

Hyaluronic acid is a polysaccharide that is naturally found in a wide variety of tissues in the human body, including the skin, eyes, and connective tissue. It's a chief component of the extracellular matrix. It attracts water molecules to itself, creating volume that can provide structural support. In a pure form, it has no tissue or even species specificity and therefore is considered to have little risk of sparking immune responses.

As such, hyaluronic acid gel fillers are used in a variety of medical procedures, with approval from the Food and Drug Administration. Hyaluronic acid (HA) fillers are injected into joints, particularly knees, to relieve pain from mild to moderate arthritis, which can decrease the natural amount of HA in joints. Age also decreases natural levels of HA, and one of the main uses of HA fillers is for cosmetic purposes—plumping lips and cheeks, and minimizing the appearance of wrinkles and fine lines in the face. HA fillers can also be used inside the eye in a variety of surgeries, including cataract extraction and corneal transplants. It can also be used topically for wound care and to relieve skin pain and itching.

For these purposes, the most common adverse effects are pain, bruising, redness, itching, and swelling, which usually last for just a few days. In extremely rare cases, there can be more serious side effects from injections, such as bacterial infections, tissue death (from blocked blood flow), and a granulomatous foreign body reaction, in which the immune system tries to clear a foreign substance, such as bacterial impurities, leading to a collection of immune cells.

A notable feature of HA fillers is that they have something of an antidote. For procedures that go wrong for whatever reason, clinicians can use hyaluronidase, an enzyme that breaks down hyaluronic acid, to dissolve the filler. Hyaluronidase is then rapidly deactivated and degraded in the body. But even without any dissolving, HA fillers offer only temporary effects, lasting between eight weeks and up to six months, for the various established uses.

Plumping penises

What is not on the list of established clinical purposes for HA fillers is penis augmentation. But that's not from a lack of trying. Studies have looked at the possibility of using HA fillers for adding girth and length—and some of have reported positive results. But so far, there's a lack of evidence to support safety and efficacy. Still, that hasn't stopped cosmetic clinics around the world from offering it anyway.

In a 2022 study in the Asian Journal of Andrology, researchers in China followed 38 patients who had HA filler injections, which seemed to provide desired results. One month after injections, the men saw an average increase of 3.4 cm in girth and 2.5 cm in length when flaccid. After a year, the measurements had decreased, but were still an increase of 2.4 cm in girth and 1.65 cm in length. Erect girth ended at an average increase of 0.8 cm. The researchers reported that only three patients had complications: two had swelling, and one had bleeding under the skin. All three cases resolved on their own.

However, not all reports are so rosy. In a 2021 case study in BMC Urology, researchers in Australia reported the experience of a 31-year-old man who received penis HA fillers at a cosmetic clinic. He developed a severe penis infection soon after that led to sepsis and multi-organ failure, landing him in the intensive care unit. Doctors ended up surgically removing pus-stained filler from his penis, which laboratory tests found teeming with Streptococcus pyogenes.

Another case study in 2021, published in a Japanese urology journal, reported that a 65-year-old man had to have part of his penis surgically removed after having HA fillers injected into the head of his penis. When he arrived at the hospital, black necrotic lesions and ulcers were readily visible.

Paraffin

While things can go wrong with HA fillers, despite their legitimate clinical uses, the outcomes for paraffin injections are much darker. Amid the Olympics scandal, most of the focus has been on hyaluronic acid. But the initial Bild article introduced speculation that skiers could also use paraffin injections to jack up their junk, which would be a terrible idea.

The use of injections of mineral oil or paraffin wax for cosmetic procedures dates back to at least 1899, when it was almost immediately found to have horrifying results. The injections lead to what's called paraffinoma, in which the body tries to encapsulate and sequester the oil it can't degrade. The result is tissue that looks like swiss cheese—with large spaces filled with oil. Over time, the tissue becomes thickened and scarred and can block lymphatic drainage. From the outside, it results in disfiguring and debilitating lesions. Some people initially ignored these disturbing results because, in some cases, the lesions can take years to develop.

While such injections have largely been abandoned, cases occasionally crop up. That includes a 2002 report of a 64-year-old Michigan man who self-injected his penis with mineral oil to increase its girth. He used a series of shots over the course of 18 months. Two years after the last shot, he went to urologists because of an increasing mass in his penis that was causing erectile dysfunction and making it difficult to urinate. He reluctantly admitted to the oil injections only after his doctors suspected cancer. In the end, the doctors had to surgically remove the mass of oily nodules from all around his penis and circumcise him.

Whatever is going on in the trousers of Olympic ski jumpers, let's hope it doesn't involve paraffin.

Read full article

Comments

Amid a push toward AI agents, with both Anthropic and OpenAI shipping multi-agent tools this week, Anthropic is more than ready to show off some of its more daring AI coding experiments. But as usual with claims of AI-related achievement, you'll find some key caveats ahead.

On Thursday, Anthropic researcher Nicholas Carlini published a blog post describing how he set 16 instances of the company's Claude Opus 4.6 AI model loose on a shared codebase with minimal supervision, tasking them with building a C compiler from scratch.

Over two weeks and nearly 2,000 Claude Code sessions costing about $20,000 in API fees, the AI model agents reportedly produced a 100,000-line Rust-based compiler capable of building a bootable Linux 6.9 kernel on x86, ARM, and RISC-V architectures.

Carlini, a research scientist on Anthropic's Safeguards team who previously spent seven years at Google Brain and DeepMind, used a new feature launched with Claude Opus 4.6 called "agent teams." In practice, each Claude instance ran inside its own Docker container, cloning a shared Git repository, claiming tasks by writing lock files, then pushing completed code back upstream. No orchestration agent directed traffic. Each instance independently identified whatever problem seemed most obvious to work on next and started solving it. When merge conflicts arose, the AI model instances resolved them on their own.

The resulting compiler, which Anthropic has released on GitHub, can compile a range of major open source projects, including PostgreSQL, SQLite, Redis, FFmpeg, and QEMU. It achieved a 99 percent pass rate on the GCC torture test suite and, in what Carlini called "the developer's ultimate litmus test," compiled and ran Doom.

It's worth noting that a C compiler is a near-ideal task for semi-autonomous AI model coding: The specification is decades old and well-defined, comprehensive test suites already exist, and there's a known-good reference compiler to check against. Most real-world software projects have none of these advantages. The hard part of most development isn't writing code that passes tests; it's figuring out what the tests should be in the first place.

The compiler also has clear limitations that Carlini was upfront about. It lacks a 16-bit x86 backend needed to boot Linux from real mode, so it calls out to GCC for that step. Its own assembler and linker remain buggy. Even with all optimizations enabled, it produces less efficient code than GCC running with all optimizations disabled. And the Rust code quality, while functional, does not approach what an expert Rust programmer would produce. "The resulting compiler has nearly reached the limits of Opus's abilities," Carlini wrote. "I tried (hard!) to fix several of the above limitations but wasn't fully successful. New features and bugfixes frequently broke existing functionality."

Those limitations may actually be more informative than the successes. Carlini reports that toward the end of the project, fixing bugs and adding features "frequently broke existing functionality," a pattern familiar to anyone who has watched a codebase grow beyond the point where any contributor fully understands it.

And that limitation is even more common when dealing with AI coding agents, which lose coherence over time. The model hit this wall at around 100,000 lines, which suggests a practical ceiling for autonomous agentic coding, at least with current models.

The human work behind the automation

Anthropic describes the compiler as a "clean-room implementation" because the agents had no Internet access during development. But that framing is somewhat misleading. The underlying model was trained on enormous quantities of publicly available source code, almost certainly including GCC, Clang, and numerous smaller C compilers. In traditional software development, "clean room" specifically means the implementers have never seen the original code. By that standard, this isn't one.

On Hacker News, the distinction drew sharp debate, reflective of a controversial reception to the news among developers. "It was rather a brute force attempt to decompress fuzzily stored knowledge contained within the network," wrote one commenter.

The $20,000 figure also deserves some context. That number covers only API token costs and excludes the billions spent training the model, the human labor Carlini invested in building the scaffolding, and the decades of work by compiler engineers who created the test suites and reference implementations that made the project possible.

And that scaffolding was not trivial, which makes any claim of "autonomous" work on the C compiler among the AI agents dubious. While the headline result is a compiler written without human pair-programming, much of the real work that made the project function involved designing the environment around the AI model agents rather than writing compiler code directly. Carlini spent considerable effort building test harnesses, continuous integration pipelines, and feedback systems tuned for the specific ways language models fail.

He found, for example, that verbose test output polluted the model's context window, causing it to lose track of what it was doing. To address this, Carlini designed test runners that printed only a few summary lines and logged details to separate files.

He also found that Claude has no sense of time and will spend hours running tests without making progress, so he built a fast mode that samples only 1 percent to 10 percent of test cases. When all 16 agents got stuck trying to fix the same Linux kernel bug simultaneously, he used GCC as a reference oracle, randomly compiling most kernel files with GCC and only a subset with Claude's compiler, so each agent could work on different bugs in different files.

"Claude will work autonomously to solve whatever problem I give it," Carlini wrote. "So it's important that the task verifier is nearly perfect, otherwise Claude will solve the wrong problem."

None of this should obscure what the project actually demonstrates. A year ago, no language model could have produced anything close to a functional multi-architecture compiler, even with this kind of babysitting and an unlimited budget. The methodology of parallel agents coordinating through Git with minimal human supervision is novel, and the engineering tricks Carlini developed to keep the agents productive (context-aware test output, time-boxing, the GCC oracle for parallelization) could potentially represent useful contributions to the wider use of agentic software development tools.

Carlini himself acknowledged feeling conflicted about his own results. "Building this compiler has been some of the most fun I've had recently, but I did not expect this to be anywhere near possible so early in 2026," he wrote. He also raised concerns rooted in his previous career in penetration testing, noting that "the thought of programmers deploying software they've never personally verified is a real concern."

Read full article

Comments

A new AMC documentary limited series on the San Francisco 49ers premieres with the first two episodes on Sunday and two more on Monday.