OpenAI this week published more than 700 AI-generated solutions to open math problems, and leading mathematicians say the work does not yet meet the field's standards. On October 6 the company pushed 722 manuscripts to GitHub, grouped into 372 families of results, all produced by an internal model it has not released. Few show their reasoning, and most were never formally checked.
The timing matters. Weeks earlier, an independent advisory group at Princeton's Institute for Advanced Study had set out recommendations for how AI labs should present math results. OpenAI helped prompt that group's creation. Yet the release appears to break its first request: to stop testing hard open problems on proprietary, unreleased models. OpenAI's own documentation says it did exactly that.
The scale of the claims is large. The manuscripts, drawn from a sample of about 4,000 open problems, include attempts at famous targets: a weaker version of the Riemann hypothesis, the Unique Games Conjecture, and the Hodge conjecture for a class of shapes. Each solution took the model about three hours of computing on average. OpenAI says it consulted prominent mathematicians first, hoping to avoid a repeat of an earlier fight over a claimed Navier-Stokes result.
The gap between claim and proof is where the field is pushing back. By TechCrunch's count, only 10 of the manuscripts included the model's chain of reasoning, and 42% were verified in Lean, a language that checks proofs line by line. That leaves almost all of them without the model’s reasoning for a human to follow, and a sizable share without a machine check either. At least three have already been pulled for errors. A paper from researchers at Cambridge and King's College London found that OpenAI's plain-language proofs and their Lean translations do not always match, casting doubt on whether a model can formalize its own work.
"When a model is prompted to solve a hard problem and spits out a solution, there is not human understanding of them at the point of release," Melanie Matchett Wood, a Harvard mathematics professor and a member of the advisory group, told TechCrunch. Terence Tao, the UCLA mathematician who announced the group, warned that "problems are being solved autonomously by AI prompters who have no interest in the broader field." James Maynard, an Oxford mathematician, told NPR that "so far it's been very difficult to really extract any human understanding from this new AI proof."
Not everyone is alarmed. "My view is that this is great for mathematics," Dan Litt, a math professor at the University of Toronto, told Fortune. The community is split, not united in rejection.
Beyond the dispute over counts and checks sits a deeper question about what a proof is for. A proof is an argument one mathematician writes so another can follow it and learn why something is true. A dense output no human can parse breaks that purpose, even if a computer confirms the logic. The advisory group framed the test as social, not technical, saying it is "ultimately up to the mathematical community to assess the extent to which our recommendations were followed successfully."
That verdict will take months. Mathematicians, not OpenAI, will decide which of the 722 files hold up, and the company has promised to fund workshops so outsiders can work through them. Until then the headline results stay unconfirmed. The real contest now is less about whether machines can solve hard problems than whether the field will accept answers it cannot yet understand.
