Nine mathematicians have agreed to advise OpenAI on how it talks about its own math research, and the company says they can publicly criticize it whenever they choose. OpenAI announced the new Advisory Group on Mathematics and Artificial Intelligence on September 21, hosted at the Institute for Advanced Study, with members including Timothy Gowers, Edward Witten, Martin Hairer and Ravi Vakil.
The group did not emerge in a vacuum. Mathematicians had already circulated an open letter titled “A Severe Misalignment of AI in Mathematics,” warning that treating unsolved problems as benchmarks for AI systems carries negative externalities for the field. OpenAI’s own post cites that letter directly and says the criticism “highlight[s] the need for thoughtful engagement between AI companies and the math community.”
That is a notable admission from a company that, in the same post, claims an internal model trained since August 28 has resolved the Navier-Stokes Millennium Prize problem and, according to OpenAI, “more than 100 long-standing open problems across most areas of mathematics.” Every figure in that claim comes from OpenAI’s own account. No outside mathematician has independently verified the count, and the post itself says the pace “surprised the mathematicians inside the company,” which is a strange thing to write if the results had already been checked by people outside it.
The group’s mandate is narrower than it sounds. OpenAI has assigned it four jobs: judge how new results get reviewed and communicated, weigh their significance, keep disclosure paced to the norms mathematicians expect, and shape how the company’s tools get used in research and teaching. One task is missing from that list. The mathematicians will not get a say in how fast OpenAI runs its internal math program, which is precisely the complaint at the center of the open letter: that speed itself, not just disclosure, is the problem.
On paper, the independence terms are real, if a little scattered. No member draws a paycheck from OpenAI, the group can rewrite its own membership, and it can publish findings the company never asked for. It can also speak up when OpenAI’s work on math falls short, in public, without clearing that criticism through the company first. “Its value depends on its members being able to exercise their own judgement and challenge ours,” OpenAI writes. Whether that holds depends on what happens the first time the group’s advice collides with a launch OpenAI has already scheduled, a scenario the post never addresses.
This follows a separate OpenAI disclosure from two days earlier, in which Noam Brown described 10,000 AI agents working on the Navier-Stokes problem in an interview. That claim and this one describe the same underlying research effort but were surfaced through different channels: one an interview, one a company blog post announcing governance around the work rather than the work itself. Readers should not conflate the two as independent confirmations of the same result.
For a math department or research lab deciding whether to treat OpenAI’s problem-solving claims as citable, the operative fact is not the count of 100-plus problems. It is that OpenAI has now conceded, in writing, that its own pace of disclosure needed outside referees before the field would trust it, and named nine people whose job is to say so publicly if it does not follow through.
OpenAI, “Advisory Group on Mathematics and Artificial Intelligence,” published September 21, 2026.
