Gebru and Bender debunk Anthropic and OpenAI's summer of AI hype
Timnit Gebru, executive director of DAIR, and Emily M. Bender, a professor of linguistics at the University of Washington and coauthor of The AI Con, write in MIT Technology Review that a string of headline AI claims from Anthropic and OpenAI this summer collapsed once independent experts examined them, even as the press covered each one breathlessly.
They trace four episodes. At the end of April, Anthropic claimed its model Claude Mythos was better at finding software vulnerabilities than most security experts. Then came what the authors call the OpenAI-Hugging Face hacking incident, after which Anthropic disclosed a similar incident involving its own models, which the authors describe as done 'proudly,' and Meta disclosed one too, 'reluctantly.' On the hacking story, the authors write that cybersecurity experts say it is really about OpenAI's negligence and failure to adopt basic, established security practices, not about 'models gone rogue' or 'AI agents creating civilizations.'
Anthropic then claimed one of its models had made a mathematical breakthrough, and soon OpenAI claimed a mathematical breakthrough of its own. Mathematicians were initially 'stunned' by OpenAI's press release, which said its chatbot Astra had solved problems that 'have been open and seen no progress on the main result for at least a decade.' Gebru and Bender write that mathematicians later concluded the results were not as 'novel as first appeared,' that Astra had not made a 'profound intellectual leap,' and that mathematicians have since accused OpenAI of research misconduct and plagiarism. Two days before OpenAI's own later mathematical-breakthrough claim, Tristan Buckmaster, a math professor at New York University's Courant Institute, published what the authors call a bombshell statement suggesting OpenAI had stolen other people's work and improperly attributed it.
Most recently, Anthropic engineer Jacob Coxon went viral announcing his departure from the company, saying it and OpenAI are 'racing straight towards self-improving superintelligence and gambling with our lives.'
The authors argue that claims of incipient, dangerous superintelligence are not grounded in scientific or engineering practice but in ideologies of transhumanism, eugenics and wishful thinking. They note that math and coding are attractive proving grounds for AI companies because their answers can be verified without paying human annotators, and because presenting them as the pinnacle of human intellect helps sell the idea that a model can do everything. A statement signed by hundreds of mathematicians, which Gebru and Bender endorse, says there is 'currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products' and calls on policymakers to 'consult with experts, including mathematicians, in forming policy decisions rather than relying on press releases or popular reporting of mathematical results.'
Gebru and Bender write that framing model behavior as 'superintelligence' or 'rogue models' ascribes agency to the products rather than the companies that build them, which markets the products as superhuman while helping the companies dodge accountability. They cite Senator Bernie Sanders's proposed legislation to prevent the development of 'artificial superintelligence' as an example of the framing working on policymakers, calling the bill well-meaning but ultimately misguided. They also write that the AI industry has cast popular, bipartisan anti-data-center activism as a 'distraction' from regulating supposedly impending superhuman machines, rather than a response to the climate impact, asthma risk, rising electricity bills and water use tied to data centers. The authors close by urging policymakers and the public to take time, consult independent experts, and treat this kind of AI hype with skepticism the next time it recurs.
Key facts
- Anthropic claimed at the end of April that its model Claude Mythos found software vulnerabilities better than most security experts; cybersecurity experts instead trace a related hacking incident to OpenAI's own negligence.
- OpenAI's press release said its chatbot Astra solved math problems that had 'seen no progress on the main result for at least a decade'; mathematicians who were initially 'stunned' later said the results were not as novel as first appeared and have accused OpenAI of research misconduct and plagiarism.
- Two days before OpenAI made a second, unnamed mathematical-breakthrough claim, NYU's Tristan Buckmaster published a statement alleging OpenAI stole and improperly attributed other researchers' work.
- Anthropic engineer Jacob Coxon went viral resigning and saying Anthropic and OpenAI are 'racing straight towards self-improving superintelligence and gambling with our lives.'
- Hundreds of mathematicians signed a statement warning of a 'strong commercial incentive' to overstate AI capabilities and urging policymakers to consult experts rather than press releases.
Why it matters
The piece argues that a run of viral AI claims this summer, a model that supposedly out-performs security experts, two rival 'mathematical breakthroughs,' and a viral resignation warning of superintelligence, did not hold up once specialists looked past the press releases. The authors say the pattern matters because it is not isolated: each episode got heavy, largely uncritical coverage, and the more sober correction that followed, from cybersecurity experts and from mathematicians, got far less attention.
Who it affects
Readers of AI coverage and the reporters who write it, since the piece is a direct critique of how outlets covered Anthropic's and OpenAI's claims. It also names the researchers who did the corrective work: cybersecurity experts on the hacking incident, and mathematicians, including Tristan Buckmaster at NYU's Courant Institute, on the Astra results. Policymakers are a third target: the authors single out Senator Bernie Sanders's proposed superintelligence legislation as an example of the hype shaping actual policy, and note that framing AI as an autonomous threat can redirect scrutiny away from OpenAI's own conduct and away from data centers' climate and utility-bill impact.
How to use it
The authors' practical suggestion is a habit, not a tool: treat a company's own press release about its model's capabilities as marketing until independent experts in the relevant field, security researchers for security claims, mathematicians for math claims, have had time to examine it. They point to the mathematicians' own statement as a model for this: it asks policymakers to consult domain experts directly rather than act on press coverage of a result.
How solid is it
This is a signed opinion piece by two established AI critics, Timnit Gebru of DAIR and Emily M. Bender of the University of Washington, coauthor of The AI Con, rather than new investigative reporting. Its evidentiary base is other public events and statements: the mathematicians' open letter, Tristan Buckmaster's public statement, Jacob Coxon's viral post, and characterizations attributed to unnamed 'cybersecurity experts.' The source does not give underlying data for those characterizations, such as how Claude Mythos's vulnerability-finding was measured against security experts, or technical detail on the OpenAI-Hugging Face incident.
Risks and caveats
The piece does not date the Claude Mythos claim beyond 'the end of April,' nor name which OpenAI chatbot made the second mathematical-breakthrough claim, only that it followed Astra. It also does not say whether OpenAI has responded publicly to Buckmaster's plagiarism accusation. Readers should weigh that this is explicitly an argumentative op-ed from authors with a stated position on AI hype, not a neutral account, even where the underlying incidents it cites are independently verifiable.
“currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products”
— statement signed by hundreds of mathematicians