OpenAI's 722 AI-generated math proofs spark a mathematician backlash

On October 6, 2026, OpenAI published 722 mathematical manuscripts generated by an unreleased internal frontier model, organized into 372 result families and posted to a public GitHub repository under an Apache-2.0 license ^[1]. The model had been posed roughly 4,000 open problems, with each published result representing on average about three hours of ChatGPT Pro-level thinking compute ^[1]. Among the headline claims: a partial 'quasi-Riemann hypothesis' result, a Lean-verified proof connected to the Unique Games Conjecture, and a resolution of the Hodge conjecture for CM abelian varieties ^[2].

But the number that matters more than 722 is 42 percent - the share of the released proofs that had actually undergone Lean formal verification, the automated proof-checking standard the field increasingly treats as a baseline for trust ^[3]. Just 10 of the resulting 719 manuscripts included released chain-of-thought summaries showing how the model reached its answers ^[3]. That gap turned out not to be theoretical: one day after release, OpenAI withdrew three manuscripts tied to the Hodge conjecture after discovering a sign error that invalidated an argument the other two withdrawn papers depended on, and separately revised 14 other papers and updated 13 citations ^[4]. The pattern - publish at scale first, discover the error afterward - is exactly the failure mode formal verification exists to catch before publication, not after it.

The verification gap compounds a second, arguably more consequential one: disclosure. In September, following an earlier controversy over claims that its models had solved more than 100 longstanding math and theoretical computer science problems, OpenAI announced the formation of the nine-member Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted at the Institute for Advanced Study in Princeton, to help rebuild trust before any large-scale release ^[5]. A week later, AGMAI issued release guidelines recommending that frontier labs disclose the specific model used, the prompts given, a summarized chain of thought, time taken, and compute cost for every claimed result ^[6].

OpenAI's October 6 release did not meet that bar. The company kept the identity of the model that produced all 722 (later 719) manuscripts undisclosed, publishing them through a company-controlled repository with only selective reasoning summaries ^[7]. University of Toronto mathematician Daniel Litt argued there is no justification for the secrecy: 'If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret from us' ^[7]. MIT's Andrew Sutherland went further, arguing that without model release and independent replication, 'you should treat any claims about one-shotting problems with a single agent as unverified' ^[6]. OpenAI effectively wrote the rulebook mathematicians are now holding it to - and missed its own bar on the two dimensions, model identity and reasoning transparency, the advisory group called out specifically.

The release's impact wasn't confined to retrospective grading of OpenAI's own rigor - it reshaped how human mathematicians were working in real time. As OpenAI's model closed in on a result related to the Unique Games Conjecture, MIT's Dor Minzer and his students Yumou Fei and Shuo Wang rushed to finish and publish their own related result rather than risk being overtaken by the company's release schedule ^[8]. Minzer described the asymmetry bluntly: 'You are human, right? You need to sleep, you need to eat, you have moods. You don't know if you're going to get scooped by the trillion-dollar company' ^[8].

Princeton's Mark Braverman framed the broader shift this competitive dynamic represents: 'Math by press release is not that healthy for math' ^[8]. The concern isn't only about credit - it's that a field's incentive structure, built around peer review, slow verification, and shared discovery, is being bent by a company that can post hundreds of claimed results at once and let the community sort out later which ones hold up. For early-career researchers especially, a problem they'd been working toward for months can be rendered moot overnight, regardless of whether the AI-generated version eventually survives scrutiny.

The sharpest institutional response came from the Association for Human Mathematics (AHM), whose statement - shared via Fields Medalist Terence Tao's blog - called the release 'not a demonstration of scholarship, but a demonstration of power' and urged mathematicians to discontinue collaboration with OpenAI ^[9]. Tao's own critique goes further than questions of verification or credit: he argues industrial-scale AI solving of open problems permanently changes the character of a field, because 'once a problem is considered solved, it can't be made unsolved again' ^[10]. In this framing, mass-producing proofs for hundreds of open problems in one release doesn't just risk errors - it strip-mines the research value out of those problems for good, foreclosing the years of community exploration, false starts, and cross-pollination that typically accompany a hard result finally falling.

That's the real stakes of the boycott call: not a one-time dispute over a sign error, but a disagreement over whether mathematics should optimize for solved problems at all, or for the slower process - what Tao calls 'Math 2.0' - of exposition, teaching, and communal understanding that a single company's release schedule cannot replicate no matter how fast its model runs.

Beneath the formal advisory critique, a more personal dispute has been driving reaction across social platforms. NYU mathematician Tristan Buckmaster has said OpenAI's release has 'destroyed the careers of early career mathematicians' - a framing that resonated widely with online commentators treating the drop less as a research contribution than as a devaluation of the work young mathematicians depend on to build a career.

That sentiment is tangled up with an earlier, unresolved dispute in which Buckmaster and Anthropic mathematician Levent Alpoge alleged that OpenAI's model reproduced a novel, unpublished approach to the Navier-Stokes equations the pair had spent roughly a year feeding into the company's tools, and that OpenAI pushed back on their authorship claims rather than crediting them; OpenAI has denied using their session data. Broader online discussion of the 722-manuscript release has skewed in a similar skeptical register - less focused on any single proof's correctness than on the optics of posting hundreds of unvetted results at once without the kind of contextual commentary a journal submission would require.