OpenAI hands maths review to an independent advisory group

OpenAI has put its mathematics claims in front of outsiders. On 21 September the company set out an independent Advisory Group on Mathematics and Artificial Intelligence: nine mathematicians who will advise on how emerging results from its models are reviewed, checked and communicated to the wider mathematical community.

The group is not a review board and holds no veto. OpenAI's own description of the arrangement is that the mathematicians established the group themselves and that it acts as a bridge between the company and the field, giving mathematicians a voice in what happens next.

What prompted it

The backdrop is a claim OpenAI has been making for weeks. The company says a new internal model, training since 28 August, resolved the Navier-Stokes Millennium Prize problem and has since worked through more than 100 long-standing open problems across most areas of mathematics. OpenAI says the pace of that progress surprised the mathematicians inside the company and prompted internal debate about how to tell the community what is coming.

That sits alongside an open letter, A Severe Misalignment of AI in Mathematics, in which mathematicians argue that treating the solution of open problems as a benchmark for new AI systems carries costs for the field. OpenAI's post acknowledges the criticism directly and frames the advisory group as its answer to the demand for engagement.

What the group actually does

Its remit is review and communication rather than discovery. The members will help OpenAI assess how significant a given result is, advise on how to coordinate its release, and set expectations around academic and professional standards of mathematical research. A second strand covers how OpenAI's tools can be used in mathematical research and teaching, which the company describes as putting capable tools in mathematicians' hands.

The independence terms, in OpenAI's words

The group will operate independently, with the freedom to offer advice OpenAI has not requested, to comment publicly on the company's impact on mathematics, and to publish its advice. Its members will not be paid by OpenAI and the group can change its own membership. One limit is stated plainly: it will not advise OpenAI on how fast to push its internal mathematics research. OpenAI calls the arrangement a first step and says its value depends on members being able to challenge the company.

Who is on it

The initial members are Francois Charles of ENS-PSL, Camillo De Lellis of the Institute for Advanced Study and GSSI, Timothy Gowers of the College de France and Cambridge, Martin Hairer of EPFL and Imperial College London, Nikhil Srivastava of Berkeley and the Simons Institute, Ulrike Tillmann of Oxford and the Isaac Newton Institute, Ravi Vakil of Stanford, Edward Witten of the Institute for Advanced Study, and Melanie Matchett Wood of Harvard. The group's page is hosted by the Institute for Advanced Study.

What is still missing

There is no timetable, no commitment to publish minutes, and no statement yet from the group itself. OpenAI has not said when the first advice is due or what would happen if the group's findings conflict with a result the company has already promoted. The group's stated power is publication, which means the test is whether anything appears under its own name.

Our opinion

The interesting number in this announcement is not nine, it is the carve-out. OpenAI has given the group three specific freedoms - to advise unasked, to criticise the company in public and to publish - and then removed the one question that would have made it a regulator: how fast the company moves. That is a fair description of an advisory panel, and it is also why the panel cannot settle the argument that produced it. The mathematicians' complaint was never that AI results go unreviewed, but that solving open problems has become a marketing benchmark, which is a claim about incentives rather than arithmetic. A group that cannot touch pacing can still inflict real damage on that dynamic, because a credible referee who publishes is worth more to the field than another benchmark table, and the names here - Witten, Gowers, Hairer, Matchett Wood - are not people who lend their signatures to be ignored. The risk runs the other way. If OpenAI's claims about a model that keeps resolving long-open problems hold up, the mathematics community is about to be asked to absorb a change in how its results are produced, and the group will be judged on whether it says so first or repeats it afterwards.