OpenAI’s Math Advisory Group Adds a Mathematician Who Signed the Letter Against It

OpenAI announced on September 21 that it is forming an independent Advisory Group on Mathematics and Artificial Intelligence, an unpaid, nine-member panel of mathematicians tasked with reviewing how the company communicates its mathematical results to the outside world. The announcement doubles as a response to three weeks of escalating conflict with the mathematics community, and one detail stands out: a founding member of the group, Fields Medalist Martin Hairer, is also one of the 25 mathematicians who signed a public declaration accusing AI companies of a “severe misalignment” with the goals of mathematical research, just ten days before OpenAI named him to help referee its own claims.

A Trajectory That Outran Its Own Credibility

The advisory group did not appear out of nowhere. It follows a seven-week run of increasingly dramatic mathematics claims from OpenAI that outpaced the mathematics community’s ability, or willingness, to verify them.

On August 2, OpenAI said an internal, not-yet-released version of its Astra model had produced results on ten long-standing open problems in mathematics and theoretical computer science, each accompanied by a Lean-formalized proof. Then, on September 8, OpenAI announced a proof that the Navier-Stokes equations governing fluid motion can develop a singularity in finite time, a partial resolution of one of the Clay Mathematics Institute’s seven Millennium Prize Problems. Crucially, OpenAI’s own writeup made clear this was not Astra’s work: “To solve the Navier-Stokes problem, we used an internal model that is significantly more capable than GPT-6 Astra.” That model had begun training on August 28, and OpenAI said at the time that “this model’s training is ongoing and its performance continues to improve.”

In its September 21 announcement, OpenAI says that same still-developing model has now “resolved more than 100 long-standing open problems across most areas of mathematics,” and adds, in its own words, that “the pace of its progress in mathematics has surprised the mathematicians within OpenAI.” Whatever this model is, internally it has gone from ten proofs to more than one hundred in under two months, with no public release, no name, and no general benchmark the outside world can check its underlying capability claims against.

Three Weeks of Backlash

That pace collided with a mathematics community already uneasy about how the Navier-Stokes result had been announced. Its rollout became tangled in a credit dispute with two outside researchers working on a closely related result (a dispute covered in detail here), and it landed the same week Caltech’s student-run “Mathathon,” an AI-assisted problem-solving competition OpenAI had agreed to help fund, came under fire from working mathematicians.

On September 10, OpenAI research lead Dan Roberts said the company would withdraw its Caltech Mathathon sponsorship after an open letter signed by 771 mathematicians objected to the event. The letter did not soften its language: “To put it bluntly, AI companies are engaging in research misconduct,” it read, arguing that companies treat research mathematics as “an advertising opportunity” and produce what commentators quickly dubbed “slop mathematics,” rushed, unreviewed output that shifts verification work onto other researchers for free. Roberts did not dispute the underlying concern: “We recognize that the rapid progress of AI in mathematics is disruptive.”

The next day, a more pointed document arrived. Twenty-five Fields Medalists, spanning medal classes from Pierre Deligne’s in 1978 to Yu Deng’s this year, published “A Severe Misalignment of AI in Mathematics” on Terence Tao’s blog. Its central claim: “solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight.” Rushed, mass-produced proofs, the signatories argued, bypass the collaboration, isolation of new methods, and citation of prior work that make a mathematical result useful to anyone besides the system that generated it. The declaration has since collected more than 7,000 signatures.

The Group OpenAI Built in Response

Nine Names, Two Fields Medals

OpenAI’s new advisory group, hosted at the Institute for Advanced Study in Princeton, reads like a direct answer to that letter. Its nine founding members are François Charles (ENS-PSL), Camillo De Lellis (IAS, 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). Two hold Fields Medals in their own right: Witten won in 1990, the first physicist ever to receive the prize, and Hairer won in 2014.

OpenAI’s own description of the group’s authority reads like an attempt to head off the usual “advisory theater” criticism such bodies draw. Members “will not be paid by OpenAI,” the group “can change its membership as it sees fit,” and it “will have the freedom to offer advice we have not requested, comment on OpenAI’s impact on mathematics, and make its advice public.” Its mandate covers “the review and communication of emerging results” and “academic and professional standards of mathematical research,” a direct answer to the citation and rushed-disclosure complaints in the Fields Medalists’ letter. One line draws a hard boundary around what the group cannot touch: “the group will not be responsible for advising us on how to pace our internal progress on mathematics.” Review, yes. Brakes, no.

One Signed the Letter, One Explained Why He Didn’t

Hairer’s presence on the roster is the detail that gives the announcement its edge: a mathematician who put his name to a declaration accusing the industry of “severe misalignment” is now one of nine people OpenAI has asked to vet how it talks about its own results. Fellow member Timothy Gowers sits on the other side of that same split, and said so publicly just four days before the group’s unveiling. In a September 17 blog post titled “Why I didn’t sign the Fields medallists’ letter,” Gowers wrote that he agreed with much of the declaration but couldn’t endorse its apparent ranking of conceptual understanding above problem-solving, since he sees “a spectrum of attitudes in mathematics to the relationship between problem-solving and conceptual understanding” rather than one correct answer. He also questioned the letter’s practical bite: “I wasn’t really sure what it was demanding that isn’t happening already.” The threat he named instead wasn’t output volume at all: it was sustaining mathematicians’ motivation and the human transmission of mathematical understanding to the next generation.

Between them, the two men represent the actual range of opinion inside the field on this question, not a hand-picked chorus of skeptics or a hand-picked chorus of enthusiasts. That range is presumably the point of putting both names on the same roster.

What the Group Can’t Fix

OpenAI has run something close to this playbook before. Twelve days earlier, it added Paul Christiano, a persistent critic of the industry’s safety practices, to the board committee that governs its safety and security decisions, using language about wanting “people prepared to challenge prevailing assumptions” on its board. The mathematics advisory group follows a similar shape: recruit a credible outside voice, give it real visibility and a public platform, and carve out the one lever, pacing, that stays under OpenAI’s own control in both cases.

That carve-out is the tension the group cannot resolve by existing. The Fields Medalists’ letter and Gowers’ dissent from it agree on more than their public disagreement suggests: what’s at stake, in both readings, is not whether OpenAI’s proofs check out, but whether a field built on slow collaborative understanding can absorb results arriving faster than mathematicians can read, verify, and build on them. An advisory group with the freedom to criticize OpenAI in public, but no say over how fast OpenAI’s own unreleased model keeps producing proofs, can referee how the results get communicated. It has no mandate to referee how fast they keep arriving, and OpenAI’s own announcement admits as much before it even lists the group’s members: “there are difficult questions ahead about how AI can support mathematical understanding and how the benefits of these capabilities can reach the wider community.”