OpenAI's new math advisory group can say anything. It can't slow anything down.

On Monday OpenAI announced the Advisory Group on Mathematics and Artificial Intelligence, an independent body of nine mathematicians hosted at the Institute for Advanced Study in Princeton. The company frames it as a bridge to the mathematical community, a way of giving mathematicians, in its own words, "a voice in how we move forward." The announcement landed ten days after twenty-five Fields Medal winners signed an open letter arguing that the way AI labs are racing through open problems is damaging their discipline.

The powers granted to the group are worth reading closely, because the interesting part is the one that was left out.

What the group can do

Members serve unpaid. They assess the significance of new results and coordinate how and when those results are released. They control their own membership, can offer advice OpenAI did not ask for, and can take disagreements public. Those are meaningful forms of independence. A body that picks its own members and can publish its objections is considerably harder to manage than a paid review panel.

Nine mathematicians were named as initial members. We are not reproducing the full roster: OpenAI's announcement page, the Institute's press release and the group's own site all refused automated access from our end, and the only member confirmed in reporting we were able to read is the Institute's Camillo De Lellis.

What it cannot do

The group will not be responsible for advising us on how to pace our internal progress on mathematics.

That sentence carves out precisely the thing the mathematicians were complaining about. The 11 September letter did not argue that AI cannot do mathematics. Its objection was that results are "announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others" — which, the signatories wrote, raises "severe attribution and plagiarism questions." The grievance is about speed and credit. The remedy announced on Monday covers assessment and communication, and explicitly does not cover speed.

The Institute for Advanced Study drew the same boundary from its own side, saying in its press release that although it will give advice, it has no decision-making power at any AI company and that responsibility for a company's decisions rests with that company.

The claim attached to the announcement

OpenAI paired the governance news with a research claim: the same internal model behind its Navier–Stokes work has, the company says, resolved more than 100 additional open problems across most areas of mathematics. That figure is OpenAI's own, published inside an announcement about how such figures ought to be communicated. We found no independent evaluation of it, and the problems and proofs are not enumerated anywhere we were able to open. It is a company statement, not an established result.

The pairing is the whole dispute in miniature. The number arrives first. The body meant to help assess numbers like it was announced in the same breath, and has no authority over when the next one arrives.

How this thread got here

Promptea covered the opening of the dispute on 8 September, when OpenAI said an AI system had produced a solution to the Navier–Stokes Millennium Prize Problem and published a Lean 4 formalization alongside it. The same day, NYU's Tristan Buckmaster announced related proofs with Anthropic's Levent Alpöge and alleged that OpenAI had acted on knowledge of their unpublished approach and pressed him over credit.

The rest compressed into two weeks. On 10 September OpenAI withdrew sponsorship of a mathematics event at CalTech after criticism from researchers there. The twenty-five-signature letter followed on 11 September, itself building on the Leiden Declaration that a working group of mathematicians released in June. Monday's advisory group is OpenAI's first institutional answer to any of it.

What actually changes

For anyone building on these models, nothing changes today. No new API surface, no model, no pricing note.

What changes is the template. Frontier labs are producing claims that the relevant expert community cannot check on the labs' timeline, and the emerging answer is a standing outside group with review duties and no operational authority. Anthropic reached a structurally similar arrangement with Accenture last week: embedded evaluators, funded by the company being evaluated. In both cases the access is real, the participation is voluntary, and the arrangement stops short of the lever the critics actually asked for.

The practical reading, for anyone who has to weigh announcements like this: a count of solved problems coming from a lab is a marketing artifact until someone outside the lab has named the problems and checked the proofs. The group announced on Monday may eventually do that work. As chartered, it cannot require that the work happen before the next number is published.

What to watch

• Whether the group publishes its own assessment of the 100-plus claim, and on what timeline.
• Whether any member resigns or goes public with an objection — the charter permits both, which makes either a real signal rather than noise.
• Whether the roster grows toward the letter's signatories, who are the constituency the group is meant to bridge to.
• Whether other labs adopt the same shape the next time a contested scientific claim comes out of a frontier model.