A serious allegation of academic intimidation now hangs over OpenAI’s claim to have solved one of mathematics’ most celebrated unsolved problems.
The ChatGPT developer announced this week that its artificial intelligence had cracked a so-called Millennium Problem, a class of mathematical puzzle that has resisted human ingenuity for decades. The specific challenge in question concerns the Euler equations, a foundational problem in fluid dynamics with roots stretching back some ninety years. The announcement was, by any measure, the kind of landmark claim that commands attention from the scientific community and from the markets that have staked considerable capital on the promise of frontier AI.
What has complicated that narrative considerably is the account offered by Professor Tristan Buckmaster of New York University, who alleges that OpenAI attempted to pressure him into collaborating on a joint announcement — and that one of the company’s scientists warned him he might “ruin his career” if he declined to co-operate. The alleged exchange, which Prof Buckmaster described in a public blog post, included the threat: “If you don’t want me to be nice, then I don’t have to be nice.” These are not the words one expects to accompany a genuine scientific breakthrough.
A Disputed Priority
Prof Buckmaster had been working on a solution to the Euler problem alongside Levent Alpöge, a researcher at Anthropic, OpenAI’s principal rival. Their approach drew on AI tools from both companies, and by the time Prof Buckmaster entered discussions with OpenAI’s scientists, he was already preparing to publish. He alleges that OpenAI’s team had only arrived at a comparable solution “in the past few days”, after rumours of his own work had circulated within the company — a timeline he finds implausible given the technical complexity of the problem. The direction of their solution, he noted, is “not the direction one arrives at in a few days.”
The question of whether OpenAI’s systems were trained on, or had access to, Prof Buckmaster’s unpublished calculations is, at present, unanswered. OpenAI’s scientists declined to address that question when he put it to them directly, which is itself a telling omission. The company has since stated publicly that neither its researchers nor its AI agents saw any of Prof Buckmaster’s work “through any means” before it was released publicly, and it has denied using identifiable user data to solve the problem. It did, however, concede that it “cannot rule out” that anonymised data were used to improve its models — a qualification that, in the circumstances, is unlikely to satisfy sceptics.
Wider Stakes for AI Credibility
The episode sits within a broader and intensifying competition among AI laboratories to demonstrate that their systems can achieve scientific results beyond the reach of unaided human expertise. That competitive pressure is entirely understandable: the commercial and reputational rewards for a verified breakthrough of this magnitude are substantial. But the pressure also creates incentives that, if poorly managed, can corrode the very standards of transparency and attribution on which scientific progress depends.
OpenAI’s public announcement on Thursday featured its latest AI system, referred to as Astra, presenting a solution to a related formulation of the problem. Whether that solution is independently valid, and whether it was genuinely reached without reference to Prof Buckmaster’s prior work, are questions that the mathematical community is now well placed — and well motivated — to examine rigorously. The integrity of the claim rests entirely on that scrutiny. A company that positions itself as a partner to science cannot simultaneously resist the transparency that science requires.