OpenAI wants to consult elite mathematicians about how to not fumble again

OpenAI announced the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), a nine-member panel of elite mathematicians hosted at Princeton’s Institute for Advanced Study. The group is supposed to advise OpenAI — and, in principle, other labs — on how AI results get reviewed, communicated, and coordinated with the wider math community.

It arrives after months in which OpenAI turned spectacular claimed mathematical results into a reputational mess: accusations of scooping researchers, weak credit for human work, and releases that fell short of mathematical practice. The same unreleased model OpenAI says resolved Navier–Stokes is now claimed to have resolved more than 100 long-standing open problems. The group’s first job is advising how to release that backlog without another fumble.

We’re glad mathematicians are in the room. We’re not yet convinced the room has power. OpenAI says AGMAI operates independently: unpaid by the company, free to offer unsolicited advice, free to comment publicly, free to change membership, with early access to research that appears necessary to do the job. Martin Hairer’s clarifying blog pushed back on misinformation, stressed genuine independence from frontier labs, and said members take representing the mathematical community’s interests seriously even while admitting nine people are not a representative sample.

Researchers talking to The Verge called it a good first step and then asked the obvious questions: how were members chosen, how much will OpenAI listen, and can an ivory-tower celebrity roster speak for working mathematicians whose careers look nothing like Fields Medal life. Simon Machado called the setup shrouded in mystery. Francesco Fournier-Facio said more engagement helps but the panel feels like an ivory tower. Kevin Buzzard wondered why OpenAI needed brilliance to hear what the community already said it wanted — better understanding of the subject, not only hard theorem trophies.

Our read: useful AI that actually solves deep math could be one of the best scientific upsides of this decade — if proofs are verified, credit is honest, and students still have a path into the field. Treating open problems as a private scoreboard hollows the pipeline that invents the next questions. Hairer’s own warning matters: labs will spin whatever the group says through a PR machine that dwarfs anything mathematicians can match.

I’m watching for the first public advice that contradicts OpenAI’s preferred timing, for a real verification pathway on the Navier–Stokes and hundred-problem claims, and for whether AGMAI becomes a template other labs copy — or a reputational shield while results drip on OpenAI’s calendar. The trajectory if this becomes normal is AI-accelerated discovery with contested credit and a profession rewriting what counts as contribution. Process decides whether that is renaissance or gutting.

The Verge by Robert Hart, Sep 23, 2026. Panel members drawn from institutions including Stanford, Harvard, Oxford, and Cambridge; hosted at IAS. Hairer’s Tuesday blog addressed origins after OpenAI approached some members and the group chose independence.