OpenAI Asked for an Advisory Board. Mathematicians Built a Tribunal Instead: Inside the New AGMAI

On September 21, 2026, one of the more unusual institutional experiments in the history of science went public. OpenAI published a blog post titled Advisory Group on Mathematics and Artificial Intelligence. Minutes later, a guest post appeared on Terence Tao’s blog announcing the same group from the other side of the table. The two announcements describe the same event in tones that reveal the whole story: OpenAI says it is “working with mathematicians who have established an independent mathematics advisory group.” The mathematicians say they were approached by OpenAI to build an external advisory board — and, “in agreement with OpenAI,” decided to build something quite different: an independent body that they control, hosted at the Institute for Advanced Study in Princeton, that answers to the discipline rather than to any company.

It is called the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), it lives at agmai.org, and its founding membership reads like a committee the field would assemble if asked to nominate its own supreme court: François Charles (ENS-PSL), Camillo De Lellis (IAS and GSSI), Timothy Gowers (Collège de France, Cambridge), Martin Hairer (EPFL, Imperial College London), Nikhil Srivastava (Berkeley, Simons Institute), Ulrike Tillmann (Oxford, INI), Ravi Vakil (Stanford), Edward Witten (IAS), and Melanie Matchett Wood (Harvard). Three of the nine — Witten, Gowers, Hairer — hold Fields Medals. conspicuously absent from the roster is Tao himself, who organized September 11’s explosive declaration by 25 Fields Medalists and lent his blog as the group’s launch platform, but does not sit on it.

Why this exists: 100 solved problems nobody has seen

The proximate cause is buried in the second sentence of OpenAI’s post, and it is easy to read past: “On August 28, we began training a new internal model. In addition to resolving the Navier–Stokes Millennium Prize problem, this model has now resolved more than 100 long-standing open problems across most areas of mathematics.”

Sit with that. OpenAI is sitting on a hoard of claimed solutions — a stockpile that, per the company’s own words, has progressed at a pace that “surprised the mathematicians within OpenAI.” The Navier–Stokes announcement on September 8 already detonated the field’s norms: a claimed Millennium Prize result released before independent verification, at a moment when two human mathematicians, Levent Alpöge and Tristan Buckmaster, were preparing related work on singularities. Three days later, 25 Fields Medalists published A Severe Misalignment of AI in Mathematics at mathandai.org, accusing the labs of treating famous open problems as free benchmark fuel — harvesting them for capability signaling while externalizing the cost of verification, exposition, and attribution onto an unprepared community.

Now imagine the next hundred such announcements queued behind that one. That is AGMAI’s actual working agenda. The group’s “current task,” per its own launch post, is “advising OpenAI on how to coordinate the release of a large number of significant results in mathematics that they report have been produced by their internal model.” And in a signal of how urgent they consider the moment, the group has opened a public input form at agmai.org/input soliciting the mathematical community’s views — with responses promised to remain confidential absent approval.

The governance design is the story

What makes AGMAI interesting is not that OpenAI sought advice — companies stand up advisory boards constantly. It is the specific inversion that occurred during negotiation. OpenAI approached some of these mathematicians about establishing an external advisory board attached to the company. The mathematicians counter-proposed, and OpenAI accepted: a group that is structurally not OpenAI’s.

The design choices, listed in both announcements, read as a checklist of every failure mode mathematicians have watched in corporate advisory bodies:

• No pay. Members do not accept compensation from OpenAI — removing the most common soft-capture mechanism.
• Unsolicited advice is allowed. The group “will have the freedom to offer advice we have not requested” and to “comment on OpenAI’s impact on mathematics” — it can critique, not just consult.
• Advice is public by default. Recommendations to AI companies will be published on the group’s own website, not delivered in private memos that die in a drawer.
• Membership is self-governed. The group can add or remove members as it sees fit; OpenAI holds no appointment power.
• Non-exclusive by construction. The group states it is “willing to offer such recommendations to any AI company whose models are likely to have a significant impact on mathematics.” It is a discipline’s interface to an industry, not a lab’s appendage.
• Powers are honestly described. “We do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company.”

That last bullet cuts both ways, and the group knows it. So does OpenAI — which was careful to carve out, in its own announcement, the one domain the advisors will not touch: “the group will not be responsible for advising us on how to pace our internal progress on mathematics.” The border of the concession is drawn with precision. Mathematicians get a voice over how results are reviewed, communicated, and disseminated; they get no voice over whether and how fast the machine that produces them runs.

Context: a week of safety choreography

AGMAI does not land in isolation. OpenAI’s publication today is one prong of a coordinated safety push ahead of Sam Altman’s in-person briefing to the UN Security Council on September 24 — an essay positioning that fully autonomous recursive self-improvement “is not happening today” and must not be pursued without safety guarantees, a call for the US to lead global AI safety standards, and now a governance structure for the single domain where its models have most visibly outrun humanity’s verification capacity. It also rhymes with the industry’s broader turn toward embedded external scrutiny: days earlier, Anthropic named Accenture’s Faculty as its first embedded evaluator, with both sides pledging at least $1 billion each over five years — advisors working inside the lab with employee-level access. AGMAI is the mirror-image model: outsiders who stay outside, publish, and owe the company nothing.

What to watch

Three open questions will determine whether AGMAI becomes a real institution or a press release. First: does the 100-problem backlog actually get released under a process the group shapes — with human verification, exposition, and attribution — or dribbled out as marketing? The mathandai.org letter’s core demand was never “stop solving”; it was “fund the human infrastructure of checking and explaining what you solve.” Second: does the group’s non-exclusivity get exercised? Google DeepMind is reportedly sitting on its own frontier math results; the moment AGMAI advises a second lab, it stops being “OpenAI’s math board” and becomes what its founders say it is. Third: what happens the first time OpenAI declines the group’s advice in public — because the group’s entire legitimacy rests on the publish-everything rule surviving the first real disagreement.

The mathematicians have, in effect, called the labs’ bluff. For a year the industry line has been that safety and stewardship require engaging experts. The experts have now shown up — unpaid, independent, and writing their terms in public. The next move belongs to the company holding a hundred unsolved problems’ worth of answers.