AI Proof Dispute: OpenAI Denies Academic Fraud Claim

OpenAI denies that its model took over two mathematicians' route to a Navier-Stokes proof, and neither side has yet produced evidence that settles it. What is on the record is two incompatible accounts and a proof no independent party has publicly checked.

What OpenAI announced

The Navier-Stokes equations describe how fluids and gases move. Since Henri Navier in 1822 and George Gabriel Stokes in 1845, it has stayed open whether a unique solution exists under all conditions. In 2000 the Clay Mathematics Institute made the question one of seven Millennium Prize Problems and attached a one million dollar award.

On Tuesday the company said a model had worked the problem through in 88 hours — on its own account running more than 10,000 AI agents, burning 130 billion tokens and millions of dollars in compute. Research chief Mark Chen described the result as a milestone.

The accusation

Tristan Buckmaster of New York University and Levent Alpöge, who works at Anthropic, went public just ahead of that announcement. They had fed their own line of attack into OpenAI tooling first, with Buckmaster using Codex, and argue the company's result rests on it. Buckmaster calls that outright academic fraud.

He also says OpenAI leaned on him once word of the work spread, pressing to drop Alpöge from the author list because of his employer, and raising the prospect of career damage. Alpöge says he overheard a hallway exchange in which a share of the Millennium prize was dangled if he stepped off the paper.

How OpenAI answers

Sébastien Bubeck says the company used neither their prompts nor their results. He rejects the idea that anyone tried to strike Alpöge from his own work, saying the dispute concerned credit on OpenAI's separate write-up. He apologizes for the career remark as a very bad choice of words.

Sam Altman confirms the company moved after rumors spread that Anthropic models had cracked a Millennium problem. OpenAI calls training on the submitted drafts unlikely, while conceding it cannot rule out that anonymized data fed back into model improvement.

Why this lands beyond one proof

Terence Tao's warning on Mastodon is about incentives, not this one paper. A rumor alone can now trigger a large AI-backed push to close a problem before the original team is ready. He expects researchers to respond by keeping work in progress to themselves, reversing what he calls centuries of open science tradition.

Outside mathematics the practical point is narrower: an opt-out from training is a promise, not an audit. No lab has publicly demonstrated that unpublished material typed into its tools stays out of the next model.

What is not established

Neither source reports a completed independent review of the proof itself. Whether OpenAI's argument holds mathematically is unsettled, as is whether the drafts entered by Buckmaster and Alpöge actually reached the model. Both points stand as assertion against assertion.