Four days after OpenAI made 722 AI-generated mathematics manuscripts public, more than 30 mathematicians interviewed by The Verge reached a similar early view: there is real substance in the release, but figuring out which results are correct may take years.
OpenAI uploaded the manuscripts to GitHub on Oct. 7. A day later, it withdrew three papers in its update log. Researchers interviewed by The Verge described the release as "astonishing," "unprecedented" and "surreal." Many said that simply understanding what OpenAI has published could itself take years, and some worried the math community may be hit with another batch before it has digested this one.
No one can yet say how many papers are correct
For now, there is no clear answer to the most basic question: how many of these papers actually hold up. OpenAI does not claim to know either. Its README says the model was evaluated on about 4,000 problems, after which the company regrouped and filtered the output into the current catalog of 719 manuscripts across 372 groups.
Those two numbers should not be used to infer a success rate. After regrouping, one problem’s output can be folded into another group.
About 42% of main results include Lean formalization
The level of verification varies widely across the archive. Lean, a programming language used to let computers check proofs step by step, offers more confidence at the logical level when a result has been formalized. In its update log, OpenAI counted 300 out of 719 main results as having formalization, roughly 42%, still less than half.
The README also says unformalized results may contain problems. Even so, interviewees said formalization is not a blanket guarantee. A computer-verified statement may not line up perfectly with the claim made in the text of a paper, so the work still has to be checked paper by paper.
Three papers have been withdrawn and 14 revised
Errors have already appeared. According to the update log, one paper on Weil classes contained a sign error that broke the argument. Two other papers built on that result, both connected to the Hodge conjecture, were withdrawn as well.
Beyond those three withdrawals, 14 papers have been revised. None of the results has gone through peer review.
Some mathematicians say the archive contains top-tier work
Set aside the way the release was handled, and several interviewees said the overall level is high. Before AI, many of these results would have been strong enough for top journals and could have helped an author secure an academic post.
Jared Duker Lichtman of Stanford University went further, saying there are "dozens" of results that would make an author a serious Fields Medal contender. His examples included progress toward the Riemann hypothesis, special cases of the Hodge conjecture and the four-dimensional Kakeya conjecture. That was his own assessment.
OpenAI said in its README that the Riemann and Hodge results were not produced through the same fixed pipeline used for the rest of the archive, but it did not explain the difference. One version of the Riemann paper also received human polishing.
Scott Armstrong of Sorbonne University noted that these are not obscure questions. Many are famous problems that generations of mathematicians have worked on for decades.
Researchers say some projects were wiped out overnight
For some mathematicians, the release was more than a shock. It meant seeing years of work lose its footing in a single night. Colva Roney-Dougal of the University of St Andrews said, "I have a bunch of friends and colleagues whose grant applications have just been wiped out."
Armstrong said he knows of one team whose research program was almost entirely cleared out. Tristan Buckmaster of New York University said he had heard of three people whose plans were erased.
The heaviest impact falls on PhD students, younger researchers and those without tenure, because unsolved problems often sit at the center of dissertations, grant applications and hiring cases. Bartosz Naskręcki of Adam Mickiewicz University said, "This is a social issue that AI labs completely ignore."
The backlash is focused less on AI doing math than on how the work was released
Most of the interviewees were not opposed to AI doing mathematics, and many said they use AI tools themselves. What has unsettled them is the method: a high-profile announcement, rough papers released into the wild, and the burden of understanding, checking and formalizing the results shifted to the wider math community before OpenAI moves on to the next target.
Among mathematicians, talk of another wave is already circulating. OpenAI did not respond to The Verge on whether more papers are coming.
Armstrong said, "Two months from now, something will come along and overshadow this." He described himself as one of the most optimistic mathematicians about AI and said he uses it too, but still admitted that he is "having a little trouble sleeping," with his immediate goal being to finish his research before "OpenAI gets there first."