On September 21, OpenAI announced the formation of an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, with professional mathematicians including Francois Charles of ENS-PSL and Camillo De Lellis of IAS. Fields medalist Terence Tao announced the news on his own blog, noting the group was created in agreement with OpenAI.
[1][2]The nut graf: after GPT-6 Astra produced a stream of math headlines — Erdős problems, a prime-gap record, and two days ago the community-exploding proof of a Liouville weak form of Goldbach — OpenAI did not keep talking to itself. It invited outside mathematicians into the review seat. But note the boundary of the arrangement: the advisory group may publicly challenge OpenAI, yet holds no power over the company's internal research pace. This is a design with a voice and no brakes.
The subtlety is in the placement. The group sits at the Institute for Advanced Study, not in OpenAI's San Francisco office; its members are professional mathematicians, not employees; its mandate is to advise on how to review and communicate emerging AI math results. In plain terms, OpenAI is conceding that its public framing of mathematical results over the past two years has damaged its credibility, and that it needs external authority to repair it. Around OpenAI's claimed 100-plus solved problems, especially the Navier-Stokes-related work, the math community has been vocal in criticizing the pattern of lab self-reporting plus media amplification. The Goldbach episode pushed that distrust to a peak: the claim came from an anonymous community account, OpenAI stayed silent, and the community could only verify via the Lean 4 files.
Read the timeline: first the anonymous claim detonated, then two days later OpenAI stood up an independent advisory group. Members may publicly challenge OpenAI but cannot stop it — an auditor shape mathematicians actually want: able to point out errors without being able to press the brakes for you. For OpenAI, this trades a slice of narrative control for the math community's trust. For mathematics, it is the first time a frontier lab has formally externalized the question of how its results get vetted.
Of course, the advisory group solves the communication problem, not the proof problem. It is not obligated to recompute every Lean project, and its advice does not constitute peer review; a claim's credibility still rests on independent verification of the proof itself. The group's real value is that "a lab says it proved a theorem" now has, for the first time, a public and accountable review process attached to it. Watch how it operates — the membership, the public statements, and the actual responses to contested results are the criteria for judging this experiment.
[1][2]