The AI summer of breakthroughs looks thinner under scrutiny

A summer of claims

The recent news cycle followed a familiar sequence:

Much of the coverage repeated the companies’ anthropomorphic language with little critical distance. Software systems became “escaped models”; ordinary product claims were framed as steps toward artificial general intelligence.

That framing turns a marketing announcement into a civilizational event before independent scrutiny has established what happened. The more useful question is not whether the language sounds dramatic, but what the underlying evidence supports.

What scrutiny found

The hacking stories were presented as admissions of failure, then amplified by the media. Cybersecurity experts, however, focused on OpenAI’s negligence and its failure to implement basic, long-established protections. Their criticism was not about models escaping control or AI agents creating civilizations.

The mathematical claims followed a similar path. OpenAI initially said Astra had solved problems that had “remained open and shown no progress on the main result for at least ten years.” Mathematicians were impressed by the announcement, but later concluded that the results were not as “new as they initially appeared.”

They then accused OpenAI of violating scientific ethics and plagiarism, while reiterating that Astra had not made a “deep intellectual leap.” Two days before OpenAI announced another mathematical breakthrough only weeks later, Tristan Buckmaster, a mathematics professor at the Courant Institute of Mathematical Sciences at New York University, published a harsh statement. He said OpenAI may have appropriated someone else’s work and misidentified its authorship.

The evidence described here does not support the idea that dangerous superintelligence is arriving through a series of clean, independently validated breakthroughs. It points instead to a familiar technology-industry pattern: ambitious claims first, qualifications and criticism later, with far less attention given to the correction.

Why mathematics and programming matter

Mathematics and programming are unusually useful showcase areas for large language models and related systems.

That combination makes these fields commercially attractive. Success in them can be presented as evidence of general intelligence, while the ability to verify answers makes system development easier.

Hundreds of mathematicians signed a statement warning that the technology industry has a strong commercial incentive to exaggerate what its products can do. They urged lawmakers to consult independent experts, including mathematicians, rather than relying on press releases and popular coverage of mathematical results.

The same statement also described an industry strategy built around speed and urgency. The authors argue that this creates the illusion that society must act immediately, diverting attention from more concrete disputes. They cite Senator Bernie Sanders’s good-faith but ultimately mistaken proposal for a bill against the development of “artificial superintelligence” as one example.

The accountability dodge

Calling products “superintelligence” or “escaped models” gives them apparent agency. That helps companies market their systems as superhuman while making it easier to avoid responsibility for what those systems do.

Instead of holding OpenAI accountable for creating malware that hacked another company, public discussion shifts toward “escaped models,” as though the software acted independently. Instead of asking researchers about companies’ alleged habit of appropriating scientists’ work or training models on user data without consent, attention moves toward imagined future machines.

My read is that the language is doing two jobs at once: it inflates the products and abstracts away the organizations behind them. The announcement cycle is notably quiet about who made the decisions, which safeguards were missing, whose work was used and who bears the cost when systems fail.

The industry has also described large, bipartisan protests against data centers as a distraction from regulating frightening “superhuman” machines. That argument asks people to worry more about an imagined machine god than about the consequences of building the infrastructure:

The sensible response is not to accept corporate urgency as a substitute for evidence. Lawmakers and local communities need time to hear independent experts and place company announcements in context. The best outcome of this period would be a pause long enough for the next spectacular claim to be recognized as marketing before it becomes policy.

Timnit Gebru is the executive director of DAIR and the author of the forthcoming book Deep Unlearning: The Radicalization of a Tech Idealist. The book is available for preorder and will be released on February 16. Emily M. Bender is a professor of linguistics at the University of Washington and co-author of The AI Con.