OpenAI Builds an Advisory Panel After a Year of Math Claims It Could Not Fully Defend

OpenAI's mathematics problem did not start with one bad tweet, though one bad tweet became its emblem. In May, VP Kevin Weil declared in a since-deleted post that GPT-5 had found solutions to 10 previously unsolved Erdős problems. Thomas Bloom, who maintains the Erdos Problems website that catalogs the open questions, called the claim a "dramatic misrepresentation," noting that several of the problems had already been solved in the literature the model was simply retrieving.

That episode set the pattern for the year that followed. In September, NYU professor Tristan Buckmaster accused the company of pressuring him not to credit a collaborator who works for Anthropic on a proof he helped verify, according to TechCrunch. Around the same time, Andreas Thom, a mathematics professor at the Technical University of Dresden, told Business Standard he suspected OpenAI's system had drawn on his own unpublished results without proper attribution. A 166-page proof the company circulated, purporting to show blowup conditions in a fluid dynamics problem, still has not been definitively reviewed by outside experts months after it appeared.

The backlash reached a formal breaking point on September 11, when 25 Fields Medal winners, including Terence Tao, Peter Scholze and Cédric Villani, signed an open letter titled "A Severe Misalignment of AI in Mathematics." The letter argued that AI labs racing to announce results on famous problems were creating negative externalities for the field, muddying credit, encouraging premature claims and forcing working mathematicians to spend their time correcting the record instead of doing research. Tao has separately warned that AI could push mathematics toward its most serious identity crisis since Gödel's incompleteness theorems.

OpenAI's response, announced Monday, is the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton. The initial roster includes François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann and Ravi Vakil, among others. Three of the nine advisers, Hairer, De Lellis and Vakil, are signatories of the Fields Medalists' letter, giving the panel a built-in line of continuity with the critics it was formed to answer.

OpenAI has been explicit about the limits of the group's authority. "It will not be responsible for advising us on how to pace our internal progress on mathematics," the company wrote in its announcement, language that leaves the panel with a voice on presentation and disclosure but none on timing or volume. With more than 100 mathematical results reportedly in the company's pipeline, according to reporting on the advisory group's mandate, the pace of output that produced this year's controversies is set to continue unchanged.

What the panel can plausibly do is narrower than what mathematicians asked for in September. The Fields Medalists' letter called for structural accountability around how results are verified and credited before release. OpenAI's framing instead treats the group as consultative, with members "free to exercise their own judgement, challenge the company" but no formal veto over what gets published or when. Some mathematicians have floated a firmer alternative: a disclosure standard requiring companies to state exactly which prior published and unpublished work a model's output drew on before any announcement goes out.

Whether the advisory group narrows that gap or simply formalizes the status quo will depend on what OpenAI does the next time a model produces a headline result. The company's own history this year, from the Erdős tweet to the Buckmaster dispute to the unresolved 166-page proof, suggests the test will come soon, and that the panel's first real assignment may already be sitting somewhere in OpenAI's pipeline of more than 100 unreleased results.