OpenAI published 719 mathematical manuscripts generated by an internal model on October 6, 2026. On October 8, TechCrunch and The Decoder both reported that the release fell short of standards set by the field's own advisory group, and some mathematicians are now urging colleagues to stop working with OpenAI. The argument is not mainly about whether the proofs are correct. It is about who pays for checking, understanding and crediting 700-plus results that nobody can simply read.
The advisory group's standards were mostly not met
Shortly before the release, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted at Princeton's Institute for Advanced Study, had published guidelines at the end of September whose first request was to stop testing advanced open problems on proprietary models — yet OpenAI used exactly that, an internal proprietary model. TechCrunch's count found that only 10 of the 719 manuscripts included summaries of the model's chain of thought. For proofs nobody can read directly, the group recommended formalization in Lean; only part of this batch received that treatment. This week, researchers at the University of Cambridge and King's College London published a paper documenting at least two discrepancies between the natural-language proof and the Lean code in an OpenAI solution to a problem derived from the Navier-Stokes equations, warning that such autoformalized proofs cannot skip peer review.
Where the boycott call comes from
A statement by the Association for Human Mathematics (AHM), shared by Terence Tao, spread quickly. It says mathematicians did not ask for this work, calls the release of more than 700 files at once a demonstration of power rather than scholarship, and urges mathematicians to discontinue working with OpenAI. Tao's own criticism is more specific: traditional breakthroughs produce talks, workshops and collaborations, whereas these results were produced by people with no interest in the field who cannot answer questions about them, leaving the work of understanding to the community. Earlier, 25 Fields Medalists had signed a statement warning that mass-producing true statements could destroy the field's fertile ground and raise attribution and citation problems. The field is divided, though: others argue public results are better than private ones and a boycott is unrealistic, and AGMAI itself is more diplomatic, insisting mainly that mathematicians must keep the freedom to choose their own questions and get equitable access to tools and compute.
Generation takes hours; understanding is billed to the community
The arithmetic explains the anger. According to complexity theorist Scott Aaronson, the results used on average about three hours of compute per problem — generation has become cheap enough for batch production. Verifying a proof and turning it into knowledge peers can use takes scarce expert time, and so far nobody is paying for it: AGMAI had suggested OpenAI help fund the human mathematicians who would have to do that work, and it did not happen with this release. What matters next is not how many more problems OpenAI can solve, but whether it adds the missing metadata, opens paths for verification, and clarifies attribution and funding. If generation costs and understanding costs stay on separate bills, the rift between mathematics and the frontier labs will only widen.
