MIT Technology Review essay: AI labs' breakthrough claims collapse under expert scrutiny

MIT Technology Review has published an opinion essay by Timnit Gebru and Emily M. Bender arguing that months of dramatic capability claims from major AI labs — spanning security, mathematics and warnings of self-improving superintelligence — have consistently fallen apart once independent domain experts had time to examine them. Gebru is executive director of DAIR; Bender is a professor of linguistics at the University of Washington and coauthor of The AI Con. Their piece, dated September 22, 2026, doubles as a media critique: the loudest coverage, they write, happens before experts weigh in, while the corrections arrive quietly.

A season of escalating claims

The essay reconstructs the sequence. At the end of April, according to the authors, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Next came a hacking incident involving OpenAI and Hugging Face, after which Anthropic and Meta disclosed similar incidents involving their own models. Anthropic then said one of its models had achieved a mathematical breakthrough, and OpenAI followed with a claim of its own: a press release stating that its chatbot Astra had solved problems that had been open, with no progress on the main result, for at least a decade. Most recently, the essay notes, Anthropic engineer Jacob Coxon went viral announcing his departure, writing that Anthropic and OpenAI are "racing straight towards self-improving superintelligence and gambling with our lives."

Gebru and Bender observe that news outlets largely echoed the companies' anthropomorphizing language, which they argue is deliberately chosen to present software as incipient artificial general intelligence.

What experts found on closer inspection

Each episode looked different under specialist review, the essay contends. Cybersecurity experts described the hacking story as being primarily about OpenAI's negligence and its failure to adopt basic, established security practices — not about "models gone rogue." Mathematicians initially stunned by the Astra release later judged the results to be less novel than presented; the essay reports accusations of research misconduct and plagiarism, and a reiterated conclusion that Astra did not make a "profound intellectual leap." Two days before OpenAI's announcement, Tristan Buckmaster of NYU's Courant Institute published a statement alleging that OpenAI had taken other people's work and attributed it improperly, according to the authors.

Why math and code are the showcase domains

The authors offer a structural explanation for why labs gravitate toward math and programming demos. Both fields are held up as the pinnacle of human intellectual achievement, which helps sell the idea of machines that can do everything — and both involve problems whose proposed answers can be verified mechanically, allowing systems to be tuned without paying data workers to annotate every output.

They cite a statement signed by hundreds of mathematicians warning of "a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products," and asking policymakers to consult experts rather than press releases or popular reporting when shaping policy.

Agency, accountability and policy

The essay's central argument concerns framing. Describing products as "superintelligence" or "rogue models," Gebru and Bender write, assigns agency to software instead of the companies that build it — marketing products as superhuman while helping vendors evade accountability. Questions about malware that hacked another company, alleged plagiarism of academic work, or the use of customer data for training without consent get displaced by speculation about future machines.

They add that superintelligence narratives rest on transhumanist ideology and wishful thinking rather than good scientific or engineering practice, and they flag Senator Bernie Sanders's proposed legislation to prevent "artificial superintelligence" as well-meaning but misguided. They also criticize the industry's suggestion that bipartisan anti-data-center activism is a "distraction," pointing instead to the climate effects of data centers, asthma in nearby communities, rising electricity bills for the public subsidizing them, and water diverted for cooling.

Why it matters

The pattern the essay documents — dramatic claim, uncritical amplification, delayed expert correction — is the practical takeaway for technical readers and policymakers alike. Verification takes time, and hype exploits that lag. Treating lab announcements as marketing until independent specialists have reviewed them, and making policy on expert timelines rather than press-release timelines, would keep attention on concrete harms — security negligence, misattributed research, data practices and environmental costs — rather than hypothetical superintelligence. The authors' closing advice is simple: pause, keep the skepticism, and recognize the pattern the next time it appears.