Mathematicians Challenge OpenAI's 166-Page Navier-Stokes Proof, Say the Method Isn't What the Company Described

OpenAI has submitted a 166-page paper that claims to resolve the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute's Millennium Prize questions. But mathematicians who have examined the work say the company does not appear to fully understand its own proof. Multiple groups of experts are disputing the claimed methodology, arguing that the approach actually used differs from the one OpenAI described in its own writeup.

The Navier-Stokes problem asks whether solutions to the equations governing fluid flow always remain smooth, or whether they can blow up in finite time. It has been open since the equations were formalized in the 19th century, and a correct proof would carry a $1 million prize. That makes any claimed solution a magnet for scrutiny, and the gap between OpenAI's description of its method and what reviewers say they found is now the center of the dispute.

The disagreement is not only about whether the proof holds. Experts say the stated derivation does not match the steps visible in the document, which leaves the actual chain of reasoning unclear. When a proof's authors cannot explain how they reached a result, the mathematical community treats that as a defect in the work, not a footnote. Reviewers are effectively being asked to reverse-engineer an argument that its own authors cannot account for.

The deeper problem is provenance. Large language models are black boxes, and the training data behind this output is not public. Even if the proof survives checking, where its underlying ideas came from will likely never be established. That raises questions about attribution and whether a result produced this way can be verified in the way mathematics normally requires, where every step traces back to stated assumptions and known theorems.

OpenAI has positioned itself as a serious player in AI-assisted science, and a Millennium Prize claim is the highest-stakes test of that pitch so far. But the episode shows the gap between generating a long, plausible document and producing a proof a community can audit. Verification, not generation, is the bottleneck, and no company controls that step.

What to watch is whether independent mathematicians produce a formal check of the argument, and whether OpenAI publishes the prompts, model version and intermediate steps behind the paper. Without that record, the dispute will stay where it is: a long document, a disputed method, and no way to tell where the ideas came from.