OpenAI's mass release of AI-generated math proofs

OpenAI's central argument for skipping peer review is that Lean formalization - machine-checkable proof code - makes manual verification 'more practical' than traditional refereeing ^[1]. In practice that substitute is incomplete and, in at least one case, misleading. As of October 7, only about 42% of the 719 surviving top-line results - 300 manuscripts - had a machine-checked Lean proof attached, meaning the majority of claims rest on OpenAI's own account of its internal model's reasoning rather than independent confirmation ^[2]. Worse, a companion analysis cross-checking one of the Lean-formalized headline results - the catalogue's Navier-Stokes manuscript - found the certificate didn't actually match the written argument: one estimate in the formalization required four derivatives where the natural-language proof claimed five, and a pressure-flux bound was proved via a different argument than the one described in prose. A Lean file that compiles cleanly, in other words, doesn't guarantee the English proof it's supposed to certify is the proof that was actually checked - a distinction the release's framing elides. Gary Marcus's critique lands on the same gap from a different direction: without disclosure of the model's architecture, failure rate, or generation procedure, there is no way to judge whether the 42% that did verify is representative of the 58% that didn't. 'This would never pass peer review,' he wrote, arguing that transparency about method, not volume of output, is what separates a scientific claim from a demonstration ^[3].

The scale of the release makes its internal dependencies hard to see, which is exactly what made the first 24 hours so instructive. On October 7, OpenAI withdrew three manuscripts - on algebraicity of Weil classes, Kuga-Satake correspondences, and the rational Hodge conjecture for products of K3 surfaces - after discovering a sign error that invalidated a stabilization-trace cancellation argument underpinning two of them. The fix rippled outward: 14 other manuscripts needed revision and 13 citations had to be updated, all within a day of a release framed as a finished, reviewed body of work ^[2]. That cascade is a preview of what happens when a single model, working from roughly 4,000 posed problems in essentially one marathon pass ^[5], builds results on top of other results it produced with the same blind spots. Andrew Sutherland's skepticism follows directly from this logic: because OpenAI's internal model remains unreleased, nobody outside the company can rerun the single prompt that supposedly produced each result, so claims of 'one-shotting' a problem are, for now, simply OpenAI's word ^[4]. The sign error shows that word isn't infallible - and raises the obvious question of how many more such errors sit undiscovered in the 58% of the catalogue that has not been formally checked at all.

OpenAI did not arrive at this release without institutional cover. On September 21, 2026 - two and a half weeks before the drop - it announced the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), a nine-member panel hosted at Princeton's Institute for Advanced Study, explicitly framed as the body guiding OpenAI's sharing practices ^[6][7]. Whatever guidance it offered didn't visibly prevent, or even shape, a single-day publication of 722 manuscripts two and a half weeks later, suggesting an advisory panel with no real veto over a lab's release calendar. The structural tension became visible within OpenAI's own circle of collaborators: Terence Tao, an AGMAI founding member, is also the mathematician who hosted and amplified the Association for Human Mathematics' statement condemning the very release his advisory panel was meant to inform, calling it 'not a demonstration of scholarship, but a demonstration of power' ^[8]. Tao's own framing goes further, describing the current moment as an unhealthy 'Math 1.0' pattern - problems solved autonomously by AI operators with no interest in the broader field - and calling instead for a 'Math 2.0' built around exposition and community rather than raw result counts ^[9]. That a sitting advisor to the company is also its most visible critic says less about any one person's inconsistency than about how far ahead of its own governance structures OpenAI's math program is moving.

The release has reshaped incentives well beyond the question of whether any individual proof holds up. Princeton's Mark Braverman put the dynamic bluntly: 'Math by press release is not that healthy for math' ^[10]. The complaint isn't abstract - mathematicians working on Unique Games Conjecture-adjacent problems rushed to publish their own results alongside or ahead of OpenAI's claims once it became clear the company was closing in, turning careful, peer-reviewed work into a race against a company's publication schedule ^[10]. Community reaction has split along similar lines: general tech audiences have gravitated toward an 'AI finally caught up to the hardest open problems' framing, while mathematicians and physicists closer to the work describe a field whose actual bottleneck is now human verification capacity rather than compute, with some PhD students reporting that in-progress thesis problems were effectively overtaken overnight. Not everyone reads the backlash as the right response, though. Dan Litt's counter-argument cuts against Tao, Marcus, and Braverman alike: '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. To me, it's going to be a good thing for mathematics' ^[9]. The disagreement isn't really about whether the proofs are correct - it's about whether a field's pace of discovery should be set by whoever can afford the most inference compute, regardless of how carefully that compute's output gets checked afterward.