"A demonstration of power": a boycott call over OpenAI's proof dump divides mathematics

On 6 October, OpenAI published a repository of more than 700 files covering 372 claimed results in mathematics and theoretical computer science, presented as solutions to long-open problems. Two days later the Association for Human Mathematics (AHM) answered with a statement reposted as a guest post on Terence Tao's blog, accusing the company of ignoring the norms of scientific research and urging mathematicians to stop working with it. Not one of those results has been independently checked. The argument that followed is not really about whether the proofs are correct — it is about who gets to set the standards of a field, and whether a field can be boycotted into agreeing.

The sentence that carried

The association's case rests on a handful of lines. "Mathematicians did not ask for this work to be done," the statement reads. Its most repeated sentence is an image rather than an argument: "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power." The text opens by noting that OpenAI "is currently defending lawsuits against accusations of illegal plagiarism, copyright infringement, and trademark dilution" — a passage that The Decoder reads as implying the results were possible only because the company trained on mathematicians' work. The press was blunter: NDTV Profit ran it as "Math Association Accuses Open AI Of Plagiarism, Copyright Violation."

The accusation of process failure is specific. OpenAI has said it consulted the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), the panel it set up at the Institute for Advanced Study. The AHM counters that the group's first advisory statement held that frontier labs should not test advanced mathematical problems on internal models, and that by ignoring that premise OpenAI showed "total disregard for the norms of scientific research." The statement closes with its demand: "We urge mathematicians to discontinue their work with OpenAI and to return to a vision of science that centers human understanding."

Who the association speaks for

This is a coalition with a view, not a parliament. The AHM is not an official body with a large membership; it speaks only for itself, and the statement is signed by its Communications Working Group. Tao, a Fields Medalist, chairs the group and gave the text his platform, but his blog labels it a guest post and adds a note that it "was initially written in a different file format and converted using AI." That caveat did not stop the statement from being read as the field's verdict. One commenter on the post asked Tao to label it "one among many perspectives" and warned that "a false narrative is now developing in many circles about how mathematicians are extremely anti-AI."

The pushback is not only about tone. It is about whether a boycott is even the right instrument for a discipline that has always asked to be judged on its arguments.

The verification problem underneath

Whatever anyone thinks of the release, somebody has to check it, and nobody has. No independent peer review of any result in the batch exists yet. Formal verification covers only part of the gap. A reader on Tao's post counted 163 of the results as shipping Lean certificates, and Scott Aaronson's own write-up is the measured one: "it also appears that no human has understood just about any of these proofs yet; the race to do so has just started." A certificate proves that a theorem follows from its assumptions. It cannot tell a reader why the result matters, which questions it opens, or which prior work it should have cited. The human referee is the bottleneck — hundreds of files, a small pool of people qualified to read them, and no institution that budgets for comprehension as a line item. Traictory's earlier report on the release found the same constraint from the other side: the binding limit is not proving, it is reading.

How much was solved, and at what cost

The scale is not small. According to The Decoder, about 8,000 problems were tested overall, with a success rate near five percent and roughly three hours of GPT-Pro-level compute per problem. Scott Aaronson called the moment a "Mathocalypse" and drew a contrast with what he calls the "Anthropic model," in which a lab works with named researchers to publish a human-readable account. The human cost is personal. Aaronson's wife, complexity theorist Dana Moshkovitz, has spent her career on the Unique Games Conjecture, which the release claims to prove. "It feels like something written by someone who's on psychedelics," she wrote, and "basically the paper is so horribly written that it's impossible to read it without AI help," describing the central construction as "some alien craziness."

The counter-argument, and the middle position

The most-upvoted dissent on Tao's post comes from a commenter writing as geeksingularity, and it attacks the AHM's premise rather than its sincerity. "Mathematics has never asked whether a result was requested. It asks whether it is correct," the comment reads, and concludes that a call to boycott is not the strong statement the field needs. On the "demonstration of power" line, the same commenter reverses the charge: "Publishing claimed proofs in full, where anyone can check them, is the most exposed position an author can take."

AGMAI occupies the ground between the two camps. In its own response it treats the release as "a first step," the beginning rather than the completion of human understanding. But it also warns that the future of the field "cannot consist only of understanding results produced by AI labs," adding that "mathematicians must be able to formulate their own questions, develop their own approaches, and explore directions that have not been selected as examples of an AI system's capabilities." That is a statement of concern without a call to walk away — the position the AHM says has already been ignored once.

What would settle it

The disagreement cannot be closed by argument, only by work that does not yet exist. Independent verification by named mathematicians, published with the errors they found rather than the verdict alone. Peer review of the results that claim to be new. Lean certificates, or human-readable write-ups, for the rest. And a ruling on the question both sides are circling: does a problem count as solved when nobody alive understands the solution? Tao has already given his answer on Mastodon, where he argues the damage is irreversible — a solved problem cannot be made unsolved, and merely knowing a solution exists can "contaminate" the search for other approaches. His prescription is a "Math 2.0" that stops treating problem-solving as the main measure of progress. He made a version of this case in September, alongside 24 other Fields Medalists who warned of a "severe misalignment" between the AI industry and mathematics, as we covered at the time. What is new is the move from warning to boycott — and the speed with which the field split over whether that move was a defense of its standards or a refusal to do the work.

The honest reading today is narrow. OpenAI released a large, unverified body of claimed mathematics. An association that does not speak for the whole field asked the field to stop working with the company that produced it. The proofs remain unchecked, the referees remain scarce, and the only thing that will settle who was right is a long, unglamorous verification effort that no one has volunteered to fund.