Announcing a result and proving it are separated by an entire discipline. Mathematics has run on its own rules for centuries: lay the proof open, let peers read it line by line, and let the next decade retrace your steps. This September, those rules met a new challenger—and received a new charter.
The Trigger: Announcing Existence Without Details
OpenAI publicly claimed its models had produced many mathematical results, including a proof of the Navier–Stokes problem—a century-old Millennium Prize Problem—reportedly achieved during a "marathon weekend of reasoning." The announcement came without details: no public proof, no way for peers to verify whether it was even correct.
On September 8, NYU mathematician Tristan Buckmaster posted an accusation that OpenAI pressured him not to mention a collaborator now employed at Anthropic in his byline, and questioned whether OpenAI had used his team's Codex-based work. On September 11, twenty-five Fields Medalists published an open letter stating that AI companies' goals were "badly misaligned" with those of the mathematical community, worrying that "the human chain of transmission between mathematicians" could be lost. The same day, OpenAI withdrew its sponsorship of a Caltech mathematics event.
Then, on September 29, the second act began.
The List of Nine
The Institute for Advanced Study (IAS) in Princeton gained a new body: the Advisory Group on Mathematics and AI (AGMA). Its nine-member list carries real weight—Edward Witten, widely regarded as physics' greatest living genius and a Fields Medalist; Martin Hairer, Fields Medalist; Timothy Gowers, Fields Medalist and the author of a long essay explaining why he had not signed the earlier open letter; plus Melanie Wood of Harvard, Ravi Vakil of Stanford, and others. All serve unpaid.
The founding method itself carries an attitude. According to AGMA's own webpage, OpenAI had approached some of these mathematicians about forming an external advisory committee. The mathematicians' response was to go build their own—independent, public, taking no corporate money. You wanted consultants; we give you a mirror.
What the Charter Says
The first output is the *Recommendations for Responsible Publication of AI-Generated Mathematics*, backed by more than 600 survey responses from the math community. The document splits cases into two tracks, with a blunt dividing line: can a human understand the proof?
If yes, follow the old road: preprints, journal peer review, academic talks—just as it has worked for centuries.
If no one can understand it, the obligations pile up. Labs must first conduct a literature search and cite everything that should be cited; the proof must be written in standard mathematical style, not output as code-generated gibberish; results must be deposited in independent academic repositories, not kept on the lab's own servers; publication must not be used as model marketing material; model names, prompts, chain-of-thought summaries, compute time, and costs must all be disclosed; results should be formalized where possible; and labs must state clearly where AI was used and how often it failed on problems of comparable scale. Beyond the checklist comes step two—labs should fund humans to understand these results: conferences, summer schools, postdocs. The money can be accepted, but decision-making authority must belong to existing non-profit organizations.
The document's hardest sentence is aimed at frontier labs: evaluating hard mathematics on proprietary models the community cannot access is a practice they do not endorse, and they ask labs to stop. The same document warns that proprietary models are splitting the global mathematical community into two tiers—those who can access the models, and those who cannot.
A Simplified Contract, and a Charter Without Teeth
For reasons of length, this article omits which specific groups disagreed in the two hundred-plus survey responses and the voting distributions across the charter's clauses; see the original document. It must also be stated plainly: where does this document's enforcement power come from? The answer is: nowhere. AGMA itself writes that it has no decision-making authority—companies remain responsible for all their own choices. Its current mandate is to advise OpenAI on how to coordinate the release of the pile of mathematical results produced by its internal model.
Can a charter with no coercive power restrain a company holding millions of dollars of compute? The mathematicians are clearly betting on that question too. Their stakes are reputation, publicity, and time. Where is this game? The charter has just been written—the match has only just begun.