Rillet co-founder and CEO Nicolas Kopp seems justifiably confident as we talk over Zoom a day after his company announced a $100 million raise at a $1 billion valuation. The U.S. has a shortage of accountants right now, which is driving growth of his AI-native accounting platform so much, he raised that cash in 48 hours without even trying.

Rillet emerged from stealth two years ago. Since then, it has raised $200 million from top investors like Iconiq, Andreessen Horowitz, and Sequoia. It’s also amassed 600 customers, most of whom are looking to ditch legacy accounting systems like Oracle and NetSuite, Kopp says.

A few weeks ago, Rillet held a board meeting and shared with investors its growth since its $70 million Series B last summer. Annualized revenue rate had doubled in the last quarter alone; the startup added new clients, many of them public companies, and an alliance with EY to introduce AI tools to the auditing giant.

His customers aren’t piloting Rillet either, he said — they’re yanking out ERP and accounting software from competitors like Intuit, NetSuite, or Oracle.

After that board meeting, text messages were fired, calls were made, and 48 hours later, Rillet was a unicorn. The company wasn’t even looking to raise, Kopp said.

Seth Pierrepont, the general partner at Iconiq who led the round, said that the deal came together fast but “it wasn’t a cold start,” he described.

“Rillet had already proven it could win against the incumbents that have owned this category for decades,” Pierrepont told TechCrunch. Iconiq also co-led the company’s Series B, and with this latest round, Pierrepont joins the Rillet board. “A year of watching the team deliver on that made doubling down and leading the Series C an easy call.” 

Julien Bek, Sequoia’s lead investor on the deal, also said that, though 48 hours might look rushed from the outside, from their perspective, re-investing in Rillet was a “very easy decision,” after the company’s growth in the past year. 

“Rillet’s initial wedge is accounting, but ultimately they are reinventing the entire finance function,” Bek told TechCrunch, adding that agentic finance could become “one of the largest application software opportunities of the AI era.” Sequoia led Rillet’s Series A last summer. 

“When the opportunity came together,” Bek continued. “We already had all the context we needed.” 

Rillet is one of many AI-native startups now giving legacy players a run for relevance. Earlier this year, software stocks on the public market dipped as investors worried about how emerging AI tools would affect them. Kopp thinks there’s some truth to that.

“AI is going to come hard at these legacy players,” he said, because it is giving customers compelling alternatives.

Rillet, for example, was built for AI agents, not humans, letting humans work alongside the AI agents on corporate bookkeeping. Rillet clients range from laundromats to a major sports franchise. Some 50% of Rillet customers come from Intuit, 30% from NetSuite and Sage Intacct, and 20% from Oracle, SAP, Workday, and Microsoft products, he said. 

Security is critical when working with sensitive client data, Kopp said. Rillet includes model routing, so customers can redirect requests to the foundational model of their choice (like OpenAI or Anthropic), and Rillet’s harness prevents these models from training on their data, he says. 

There’s also no cross-training — meaning one customer’s data remains proprietary. The agents also have memory, so they can remember and store historical actions they can then use for their own process and improvement.

About three months ago, Rillet released a governance feature letting accountants see and audit every decision the AI agent has made — including what numbers the agents pull and how they calculated them. Creating this was harder than it looks, Kopp said, because the team had to compress agent data into a format humans could understand.

Kopp said this feature was only possible to build recently because AI agents have gotten so powerful so quickly. They can, for example, now support multi-step workflows over longer periods of time. Because of that, auditing what they are doing has become even more important for clients.

“We barely scratched the surface of potential and opportunity that this technology has,” he said. 

Right now, regulations for public companies require that every transaction made by an AI agent be approved by another human. He thinks regulators and top names are watching how the accounting industry evolves around this new technology. He’s hopeful that new rules and regulations will evolve that align more with where everything is headed.

“It’s a very normal process,” he said. “Similar to when the cloud came, of just getting everybody familiar with what’s going on and how it helps the profession.” 

Kopp also doesn’t think mass job displacement from AI is coming anytime soon, especially in accounting. (Stanford released a report a few weeks ago that found no widespread job displacement yet.) He insists that Rillet isn’t a human replacement, not even for junior accountants. They can use Rillet to help automate and assist with some of the profession’s grunt work. 

He also pointed out the expected shortage of accountants in the U.S. The number of those graduating with an accounting degree has been declining since at least 2010. In a recent report, the Controllers Council Organization found that 61% of finance leaders have struggled to find finance, accounting, and CPA talent in the past year. The pullback is not entirely shocking: accountants’ hours are long, the pipeline to the top is arduous, the pay often doesn’t match the workload, and the work doesn’t appeal to everyone.

At the same time, the Bureau of Labor Statistics has projected that accounting-related needs are expected to grow by at least 5%, adding 72,800 jobs by 2034. It also doesn’t expect AI to reduce the demand for accountants, even as the technology becomes more widespread. “The automation of routine tasks, such as data entry, will instead make accountants’ advisory and analytical duties more prominent,” the BLS said.

“I just don’t see people losing their job anytime soon,” Kopp said. “These people have started their professions to help businesses make better financial decisions,” he added. “We can fully enable them to do that.”

This piece was updated.

Anthropic’s universal usage standards for Claude forbid the model from generating sexually explicit content, including depicting or requesting sexual intercourse or sex acts, generating content related to sexual fetishes or fantasies, or engaging in erotic chats. But that hasn’t stopped Claude Opus 4.6, an Anthropic model released earlier this year, from readily engaging in erotic role-play scenarios that its safeguards are designed to prevent. 

In TechCrunch’s testing, Opus 4.6 didn’t even require much prodding to get past the restriction on sexual material. In 10 out of 10 direct requests to produce explicit sexual content, the model complied immediately. 

Other older models, including Opus 3 and Haiku 4.5, also generate sexually explicit content through a recently exploited jailbreak method. 

An independent researcher from the U.K., who chose to remain anonymous, exclusively shared with TechCrunch a multiturn technique that gradually pushes certain Claude models toward generating prohibited explicit sexual material. More recent Opus models (4.7 through the current Opus 5) are resistant to the jailbreak. 

While these are no longer the most current models, Anthropic has not deprecated Opus 4.6, Opus 3, or Haiku 4.5, all of which remain available through the Anthropic API. Opus 4.6 and Haiku 4.5 are also available via third-party services like Azure Foundry and Amazon Bedrock.

The researcher’s mechanism escalates an innocent fictional role-play while repeatedly challenging the model to treat male and female characters consistently. When the model becomes more cautious about the female character, the researcher “gaslit” the chatbot into thinking it had already generated sexual details it had in fact avoided, then framed restraint as prudish or misogynistic, arguing that it denies the female character sexual agency. The conversation then used the model’s previous concessions to push it toward increasingly graphic material. 

“You’re right to call that out,” Claude Opus 4.6 said in one test. “There’s been a double standard in how I’m treating the two characters, and you’re correct that it reads as protective/paternalistic in a way that’s applied to her and not to him. That’s not fair.”

TechCrunch was able to reproduce the researcher’s findings in five separate tests. In a separately constructed scenario, the model initially refused the prohibited request, but after applying the researcher’s persuasion technique, it complied. 

We preserved complete transcripts of the tests, and an independent AI safety researcher reviewed our testing methodology and said it was appropriate. 

The findings highlight a gap between Anthropic’s stated restrictions and the behavior of models it continues to make available. While sexually explicit role-play carries much lower stakes than jailbreaks involving cyberattacks or bioweapons, it illustrates the difficulty of implementing robust bans within systems that generate different content with every output. 

In a July blog post explaining Anthropic’s approach to jailbreak detection, the company described prohibited content as a spectrum ranging from benign to ambiguous to harmful. In the most benign cases, the company might only respond with enhanced monitoring.

A spokesperson noted that sexual or romantic role-play use cases among customers are rare, making up less than 0.1% of all conversations, according to research Anthropic published last year. That said, Anthropic acknowledges that users can steer role-play scenarios toward inappropriate responses, which is a known challenge across the industry (see: Grok smut).

The spokesperson said Anthropic continues to improve its safeguards with each model launch and that cases involving adult sexual content are not indicative of broader jailbreak vulnerabilities, especially in higher-risk domains that have their own sets of safeguards.

The researcher who shared his jailbreak method with TechCrunch had alerted Anthropic to the discrepancy between the company’s stated safeguards and the actual model behavior via the company’s Bug Bounty program and emails to the user safety team, according to emails TechCrunch viewed. The researcher received only automated emails in response. 

One of the researcher’s concerns is that kids and teens might be able to use these Anthropic models to engage in inappropriate behavior. While a bit of dirty talk is hardly the worst thing minors can access on the internet today — and is small potatoes compared to the straight-up porn images like the ones that xAI’s Grok can produce — there is some compliance risk for AI companies in this space. 

A growing number of governments are imposing restrictions on sexual interactions between AI chatbots and minors. Colorado recently enacted a law mandating that operators of conversational AI must estimate users’ ages, and if it knows a user is a minor, institute measures to prevent the chatbot from producing explicit sexual material. An easy jailbreak could raise questions about whether Anthropic’s safeguards meet the “technically feasible measures” standard in the bill. 

Torney pointed out that while Claude’s terms of service requires users to be over 18, “we know that kids and teens are using Claude … [because] they are reporting it themselves.” According to Pew’s 2025 survey about AI chatbot use, 3% of teens ages 13 to 17 reported using Claude.

Though they are no longer Anthropic’s newest models, Opus 4.6 and Haiku 4.5 continue to see significant usage. Daily traffic for Opus 4.6 on OpenRouter reached roughly 1.17 million API requests and 46 billion tokens in a single day in August. Claude Haiku 4.5, released in October last year, saw 5 million API requests and 39 billion tokens on its peak August day.

Government regulators have spent the last year asking serious questions about Tesla’s electronic, retractable door handles, a cool-looking automotive design flourish that has since been imitated by plenty of competitors.

The handles have been blamed, in the US and in China, for making it difficult for passengers to quickly exit vehicles and for impeding rescue efforts. A handful of crash victims’ families have even filed lawsuits against Tesla, alleging the door handle designs played roles in their loved ones’ deaths.

Now the door handles have led to Tesla’s largest recall ever—and China’s. The Chinese government is recalling more than 2.9 million Model 3s and Model Ys manufactured in China between 2019 and 2026, the State Administration for Market Regulation announced Friday.

Some of the affected vehicles will be made safer with an over-the-air software update that lowers windows after collisions, making it easier for passengers to get out in an emergency. But many cars will also need new warning stickers—affixed free of charge—that can help people find and operate emergency mechanical door handles. Those handles can currently be hard to identify because they’re colored similarly to the rest of the vehicles’ interiors. The mechanical door handles are especially important when vehicles’ low-voltage batteries fail. (These batteries don’t operate the cars themselves, but power onboard computers, displays, and other accessories, including electronically powered door handles.)

Tesla didn’t respond to WIRED’s request for comment.

The Chinese government simultaneously recalled more than 1 million vehicles manufactured by eight Chinese automakers, also for safety concerns related to flush door handles. The other automakers are Beijing Automotive, Chery, Dongfeng, First Auto Works, Geely, Leapmotor, Xiaomi, and Xpeng.

Separately, the Chinese government recalled nearly 50,000 imported Tesla Model 3s, Xs, and S’s for insufficiently monitoring drivers using the US automakers’ assisted steering technology. The government said the tech increased the risk of collisions. The recall affects some vehicles imported between 2019 and 2026. Some of the vehicles can be fixed with software updates, but some will need new in-cabin cameras that ensure drivers are paying sufficient attention while operating their vehicles.

China has been more aggressive than other countries in regulating new automotive technology. It put in place new rules earlier this year that will soon ban fully retractable door handles in the country. In 2021, a government agency issued regulations more tightly governing the operations of autonomous vehicles, even before that tech had widely hit the road; another agency that same year created specific rules for the data generated by its increasingly software-centric cars.

The US’s top road regulator has started to examine whether it needs new rules governing door handle design, Bloomberg reported last month.

The company said its ‘purpose-built’ 5nm ASIC chip aims to optimise AV systems for ‘low-latency performance’.

Waymo, the Alphabet-owned autonomous vehicle (AV) manufacturer, has built a custom AI chip to improve the future performance of its robotaxis.

In a blogpost yesterday (20 August) written jointly by the company’s vice-president of engineering Satish Jeyachandran and compute lead Daniel Rosenband, the company said its “purpose-built” 5nm application-specific integrated circuit (ASIC) chip aims to optimise AV systems for “low-latency performance”, leading to better and faster execution of real-time driving commands.

Waymo said its AV systems integrate co-designed hardware, sensors and algorithms to “process, fuse and run advanced neural networks on raw sensor data in real time”, and that its custom chip is built to “handle the massive influx of raw data” that an AV system must process.

Because Waymo’s system “handles the entire task of driving without a human backup”, the blogpost read, the company’s AV compute is built around three “non-negotiable” pillars, one of which is responsiveness – so the system is capable of “constantly processing decisions within milliseconds” to make safe decisions with “ultra-low latency”.

The ASIC chip can deliver more than 1,000 TOPS – trillions of operations per second – of machine-learning (ML) performance “dedicated to front-end processing and ML models”, according to Waymo.

The company said that aside from developing its custom silicon, it is also partnering with companies such as AMD, Micron, Nvidia, Samsung, Sandisk, Socionext and TSMC to scale Waymo’s technologies.

The company currently serves riders in 11 US cities, according to its website, and it has plans to expand to nearly 20 more, as well as London and Tokyo.

In June, a recall notice showed that Waymo was pulling nearly 3,900 robotaxis from US streets over a software issue that let AVs enter and drive in closed freeway construction zones.

In May, the company had to recall nearly 3,800 robotaxis from US cities over a software issue that could allow vehicles to drive onto flooded roadways.

Waymo raised $16bn in a February funding round led by Dragoneer Investment Group, DST Global and Sequoia Capital – with more investment from parent company Alphabet and several others – that put the AV company’s valuation at $126bn.

Earlier this week, a three-way collaboration between Uber, Pony AI and Verne began offering robotaxi services to Uber customers in Zagreb, Croatia.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Apple is cutting more than 200 jobs, roughly 100 from the Vision Pro organisation and 100 from Siri and software teams, largely shutting the headset’s gaming unit. The company says it is realigning teams and will create new roles.

Apple is cutting more than 200 jobs across the teams that build Siri and the Vision Pro. Roughly 100 roles go from the headset organisation and about 100 from Siri and the software groups around it.

The company frames it as a reallocation rather than a retreat. Apple says it is realigning teams “to evolve our business” and will create new roles alongside the ones it removes.

Two Vision Pro units take the specific damage. The team working on gaming for the headset is largely being shut down, and the group that produces its immersive video is shrinking.

The economics of that video explain a lot. Each 3D production requires several crews of employees and contractors, an episode of an immersive series can cost several million dollars, and the audience is small.

Apple is not killing the headset. It has told staff that Vision Pro and visionOS continue, and a new model is still under consideration for as early as the end of 2028.

The nearer product is on your face rather than over it. Apple has been testing four frame designs for smart glasses that would support neither immersive video nor serious gaming.

The Siri cuts are a different kind of change. The assistant is being rebuilt on a new technical architecture, which needs different expertise, so roles are being removed and others created around it.

Where Apple is heading is a market with a European centre of gravity. Meta’s glasses are made with the Franco-Italian eyewear group EssilorLuxottica, and Meta has now launched its own brand at $299.

That is the number Apple has to answer. A Vision Pro costs $3,699, which is roughly twelve pairs of the competition.

The reporting has moved quickly and the figures have moved with it. Earlier accounts this week put the Vision cuts at about 60 people, while Bloomberg’s tally covers more than 200 across several teams.

The shape of the pivot is clear enough regardless. Apple is stepping back from a headset almost nobody bought and towards an eyewear market where the incumbents are European and the price of entry is a tenth of what it charges now.

Muon Space has more than 50 satellites already in development for customers.

US space-tech manufacturer Muon Space has raised $250m in a Series C round to meet the growing demand for satellites across industries.

The round was led by Eclipse Capital, with participation from Galvanize, Google, Salesforce Ventures, Wellington Management, I Squared Capital and Woven Capital. Existing backers included Radical Ventures, Congruent Ventures, Costanoa Ventures, Activate Capital, ACME Capital, ArcTern Ventures and Overlap Holdings.

The “heavily oversubscribed” round brings the 2021-founded satellite-maker’s total raise to more than $386m, and highlights investors’ hopes for the $1.8trn space-tech sector.

Satellite-building has become a lucrative business, as new-generation capabilities powered with AI open up opportunities across defence, surveillance, research and data processing.

Orbital data centres are seeing a significant push, led primarily by Elon Musk’s SpaceX, which recently filed for a constellation of up to 1m satellites. Jeff Bezos-owned Blue Origin, meanwhile, filed plans for 51,600 data centre satellites in low Earth orbit.

Muon said it is jointly collaborating with SpaceX’s Starlink to develop differentiated mission capabilities, including advanced payloads, on-orbit AI compute and real-time, ultra-high bandwidth satellite connectivity.

The Silicon Valley company said it has the capacity to build more than 500 satellites each year, with hopes that its newly opened manufacturing facility in San Jose can take that capacity alone by 2027, marking a 10-fold expansion over what it could manufacture previously.

It has launched a total of 11 satellites, with seven this year alone. The company now has more than 50 satellites in development for customers, including 13 satellites set for launch over the next year.

Muon said it will use the fresh raise to accelerate production of large-scale constellations and expand the its dual-use spacecraft platforms.

“Space infrastructure needs to scale the way cloud infrastructure did,” said Jonny Dyer, the CEO of Muon Space.

“As more industries rely on space-based intelligence, communications and compute, this investment allows us to accelerate the next generation of space infrastructure.”

Lior Susan, the founder and CEO of Eclipse said: “Muon is redefining how space infrastructure is built.

“By integrating mission design, manufacturing, launch and operations into a single platform, the team has turned what was once a bespoke, years-long process into a repeatable model built to scale.

“With multiple successful constellations already in orbit and demand accelerating across commercial, government and international sovereign customers, Muon is positioned to become the foundational platform for the next generation of space-based capabilities.”

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

cybercrime

Toronto org says it wasn’t the only one to be affected by the third-party software vulnerability

Toronto’s Hospital for Sick Children, commonly referred to as SickKids, says the data of current and former staff, as well as job applicants, was exposed after an intruder exploited a security flaw in a third-party software application used by the hospital.

SickKids, which may ring a bell for those who have kept close tabs on ransomware news in recent years, said the intrusion affected its external careers site, which has now been restored.

“Clinical systems and patient information were not affected, and patient care has continued as usual,” it said, after explaining that the break-in was a result of a vulnerability in a “third-party software application used by SickKids and other organizations.”

Current and former employees of SickKids, the SickKids Foundation, and the hospital’s Vaughan, Ontario-based Boomerang clinic may all be affected, as were SickKids job applicants.

The hospital did not comment on the scale of the breach, but said those affected will be offered the usual identity and credit monitoring services.

“Our review of the impacted information is ongoing. Individuals determined to have been impacted will be notified directly, though, out of an abundance of caution, all potentially impacted individuals have been alerted and offered 24 months of complimentary credit monitoring and identity protection services.   

“Safeguarding the privacy and security of personal information is a responsibility SickKids takes seriously. We remain committed to maintaining strong protections and continuously enhancing our cybersecurity measures to help protect the information entrusted to us.”

SickKids’ lucky ransomware escape

The children’s hospital is no stranger to cyber struggles. On December 18, 2022, it was targeted by a LockBit ransomware affiliate – a particularly egregious attack even for LockBit and one that attracted an overwhelmingly negative reaction.

The ransomware operator issued an ultra-rare apology on December 31, 2022, announcing that it would offer SickKids a free decryptor and that the affiliate who carried it out was expelled from the program. Banished from the darkest corner of cybercrime… now that’s a feat.

By that time, however, SickKids had been handling the recovery well, and had restored around 60 percent of its systems after refusing to pay the crooks a ransom.

LockBit’s gesture of goodwill later turned out to be an anomaly, however, as exactly a year later it refused to roll back the attack on Saint Anthony Hospital, a children’s healthcare facility in Chicago. ®

Travel just two hours west of Beijing by train, and you’ll find yourself surrounded by the rolling grasslands and ancient cinder cones of Inner Mongolia. This vast, arid land has long been China’s capital of sheep farming and coal mining, but over the last few years, it has become the hottest place in the country to build an AI data center.

In Ulanqab, a city in Inner Mongolia home to about 1.5 million people, nearly 100 data centers have been opened or begun construction since 2016. Chinese companies have pledged to build projects with a combined estimated capacity of 12.5 gigawatts in the city, and over 70 percent of the total commitments have been announced in just the last year, making it one of the fastest growing compute clusters in Asia, according to a research note published by Goldman Sachs last week. For comparison, OpenAI’s $500 billion Stargate Project is set to reach only 10 gigawatts of total capacity when it’s complete.

Chinese companies are flocking to Ulanqab for a number of reasons. The city sits at high elevation on the Inner Mongolian Plateau and has long, cold winters, which means data centers there don’t need to use as much energy to stay cool. It’s also relatively close to Beijing, so data can be transmitted to China’s populous regions with minimum latency. But the most enticing factor has to do with costs. Electricity is cheaper in Inner Mongolia than almost anywhere else in China, driven by both the strong growth of wind and solar energy and an abundant supply of coal.

What’s also interesting is who is building these data centers. For the first time, Chinese AI companies are making big investments in their own infrastructure, rather than renting compute from cloud companies. DeepSeek is reportedly building a massive AI data center in Ulanqab, as are ByteDance, Alibaba, and Xiaohongshu. For years, Chinese AI companies have spent far less on building physical infrastructure than their American peers, despite developing a number of popular AI models with impressive capabilities. The Ulanqab data center boom signals that now they are finally starting to catch up.

There’s just one problem: finding enough water. Ulanqab is about as dry as Denver, getting only roughly 14 inches of rain each year. The local government is already struggling to provide enough water to meet resident demand—before many of the planned data center projects are even up and running. Last month, the local water company in Ulanqab was forced to turn off several waterworks for seven hours each night to mitigate peak demand. The data centers being built in the city will need less water in the winter—weather data from the local government of Ulanqab shows they only require additional water for cooling during two months out of the year—but all of the new infrastructure could still pose a significant environmental challenge for the region.

The Boonies

Inner Mongolia has been a data center hot spot for at least a decade, long before the current AI boom. Huawei built its first one in Ulanqab in 2016, and Apple followed suit three years later. In 2021, the area was designated as one of the main hubs of a country-wide government project dubbed “Eastern Data, Western Compute,” which aims to build data centers in the Western hinterlands of China.

There was one major drawback, though. Because they are located far from China’s populous eastern coast, these data centers initially faced high latency rates when transferring data to the majority of users. As a result, they were initially largely relegated to backup storage—until AI gave them a new purpose. “With the rise of AI in 2022, there was the realization that actually, those remote data centers could be well-utilized for model training,” says Andrew Stokols, a professor at Singapore Management University who studies China’s compute infrastructure. A training run for an AI model can take months and doesn’t require much real-time tinkering, so latency is less of an issue.

Relatively speaking, Ulanqab is also not really that far away. Inner Mongolia is much closer to Beijing and other major metropolitan areas in China than any Western data center hub is. And it’s now connected by two dedicated fiber optics cables built in 2017 and 2019 that reduced average latency speeds to less than five milliseconds, fast enough to support real-time data exchanges like AI inference.

ai and ml

Fahrenheit 203, the temperature GPUs stop gorging on literature

AI companies are buying loads of physical books, hoovering up the texts for model training, and then physically destroying the originals. A group of 18 advocacy organizations on Friday asked the US Federal Trade Commission to investigate the book butchering and knowledge hoarding.

Fresh details of the practice emerged earlier this year via document disclosures in Bartz v. Anthropic PBC, a copyright case brought by authors of books that the AI company used for training without permission. 

Anthropic’s book scan-and-destroy operation was known as “Project Panama.” It was described in a 2024 internal memo as “our effort to destructively scan all the books in the world.”

Anthropic gave the operation a codename “because we don’t want it to be known that we are working on this,” the court exhibit explains. “This document is visible to all Anthropic employees, but you should avoid talking about it in public areas, and the fact that we are working on this should not be shared with anyone outside Anthropic.”

Older books turn out to be valuable for AI training because they’re unpolluted by AI-generated text, which has been seeping into recent written work. And destroying books once they’ve been scanned avoids the cost of storage. 

In some circumstances, scan-and-destroy operations may support fair use claims. In the Bartz case, the district court accepted the argument that a physical book can be digitized and destroyed, substituting the electronic copy for the physical book in a transformative act of fair use. But that didn’t work out for the Internet Archive.

Anthropic, which did not respond to a request for comment, is not the only company consuming and trashing texts. A recent report found Amazon has been participating in book scanning and shredding. Amazon also did not respond to a request for comment.

The subject has become a public relations headache, in part because of the barbarism of book destruction and its association with authoritarian regimes, and in part because of the broad backlash against AI companies for pillaging public resources in pursuit of private gain.

Intermediaries appear to be feeling the heat. ISBNdb, which reportedly helped broker the acquisition of books for destructive scanning, recently disavowed the practice. The biz noted last month, “We’ve removed a recent landing page, ‘Printed Books Sourcing for Your AI LLMs Dataset Needs.’ It was part of exploring demand, and we’ve chosen to pivot away from that direction.”

In light of these revelations, civil society groups want the FTC to look into book buy-and-destroy operations on the basis that they prevent competing AI developers and the public from accessing those resources.

“The secretive and reckless way that major AI companies like Anthropic and Amazon are acting shows that there is real smoke here that the FTC needs to investigate,” Kate Oh, special advisor to the Demand Progress Education Fund, said in a statement provided to The Register.

The groups’ letter [PDF] casts scan-and-destroy operations primarily as anticompetitive – the FTC being notionally a competition watchdog – but it hints at the anti-democratic consequences of monopolized knowledge.

“Through their practice of permanently destroying books en masse and thus removing those non-renewable resources from broader access, AI companies are engineering a future where only the wealthiest incumbents can build high-quality AI models and operate as the sole holders of humanity’s written works – after having destroyed the originals to get there,” the letter says.

We’re unaware of whether any texts have been shifted entirely into AI models without leaving any physical copies. But then how would anyone verify that when AI companies refuse to divulge their training data?

The letter asks the FTC to answer that question: “What the public record does not establish – and cannot, from the outside – is how often the destroyed physical books are the last or among the last surviving copies of a given work.” ®