AINEWS
Mathematicians want a verification and citation standard for proofs that come out of a model, and OpenAI says no unpublished work was used.
Mathematician Tristan Buckmaster went public on September 10 with a specific accusation: OpenAI's system had seen his group's private work on the Navier–Stokes problem before OpenAI published its own Navier–Stokes solution. The New York Times reported Buckmaster's claim. OpenAI told USA Today the accusation was "categorically false" and that no user data was accessed.
Navier–Stokes is one of seven Millennium Prize problems, the Clay Mathematics Institute's list of the hardest open questions in math, each carrying a $1 million award. A complete solution would prove that the equations describing fluid flow never blow up into infinities, settling a question mathematicians have worked on for centuries.
OpenAI said the work came from an internal system running a model "significantly more capable than GPT-6 Astra." The project started September 1, two days after the company heard rumors that two Millennium Prize problems had been solved elsewhere. The Navier–Stokes post does not name a human author and does not say how the candidate proof was checked.
AI assistants already help mathematicians in narrow ways: they suggest next steps in a stalled proof, draft candidate lemmas, and search the literature for prior results. Mathematicians have published papers that lean on those uses, and the field has not collapsed. The objection is to a different question: whether pattern-matching over human proofs counts as doing mathematics, and what verification and citation rules should apply when a model produces a candidate step.
OpenAI researcher Sebastien Bubeck called Buckmaster's allegation "false and inflammatory." Sam Altman said Bubeck "acted with integrity and generosity throughout," per ValueAdd VC's recap of Forbes coverage. Scientific American framed the episode as "research misconduct" by OpenAI. The Verge said the company had sent "a chill through academia." Futurism covered the dispute as a public-relations problem for OpenAI. Hacker News commenters criticized the Economist's coverage for ridiculing the mathematicians it reported on, with several pointing out that the original piece opens with a comparison to Socrates complaining about writing.
The factual question, whether OpenAI's system saw unpublished work, remains unresolved. OpenAI says no. Buckmaster disputes that. Without training-data logs, the verification chain used on the candidate proof, and a model card for the system that produced it, neither side can close the gap. The accusation collapses if the logs are clean and the verification chain is documented; the skepticism hardens if either is missing or contradicted.
The structural question is easier to answer. A mathematical proof is a sequence of statements in which each step follows from definitions, prior results, or earlier steps, with no gap a reader cannot fill. Mathematicians can name the rigor bar: a complete proof written in a form a competent peer can audit, step by step, without seeing the model that produced it. OpenAI can describe its workflow: which steps were model-generated, which were human-checked, what datasets the model was trained on, and which published results informed each step. The Navier–Stokes post names none of those. The post is a claim, not a proof.