OpenAI Posts 719 Math Manuscripts, and Scientists Demand They Be Checked

OpenAI has posted 719 math manuscripts on GitHub, and the professional community has answered with a list of complaints. According to TechCrunch, some of the material has not been fully vetted, and some proofs may contain errors. The company's claim to have solved one of the Millennium Problems, which involves the Navier–Stokes equations, raised separate questions.

What was published and what bothered scientists

On its face, the release looks like a show of scale: 719 texts in an open repository. Mathematicians, however, care less about volume than about the status of each document. A manuscript on GitHub is not an article in a peer-reviewed journal. It has no editor, no anonymous referees, and no mandatory discussion in the relevant research community. Anyone can upload a file, and no one is obliged to read it.

According to TechCrunch, some of the material has not yet been fully vetted. Critics allow that some of the proofs contain errors. For mathematics, this is the core objection: a proof is worth exactly as much as it can be verified. If it is unclear which of the 719 texts are correct and which are not, the collection is of little use, and the burden on experts has grown.

Key parameters of the release:

• volume: 719 manuscripts;
• venue: GitHub, meaning a preprint format without peer review;
• verification status: some of the material has not been fully vetted by experts;
• risk: individual proofs may contain errors;
• most disputed claim: a solution to a problem related to the Navier–Stokes equations.

Navier–Stokes: why this claim demands extra caution

The Navier–Stokes equations describe the motion of viscous fluids and gases. The question of whether their solutions exist and are smooth is one of the seven Millennium Problems listed by the Clay Mathematics Institute. Each carries a prize of one million dollars. Of that list, only the Poincaré conjecture has been fully resolved.

The institute has its own rules: a solution must appear in a respected peer-reviewed publication and stand up to the test of time. A press release or a repository does not meet those conditions. The scientists' reaction is therefore predictable. They expect a claim of this magnitude to arrive as a detailed paper that can be examined step by step, not as one entry in a large pile of files.

What mathematicians are asking for

The complaints boil down to three points. Scientists want full scientific papers for the results, disclosure of the methods the AI used to obtain the proofs, and a clear statement of what the researchers did and what the model did.

The second and third points are linked. If a model built a proof and humans then edited it, the reader needs to know where the line fell. Who is responsible for an error? Which steps were checked by hand, and which were taken on trust? Without that information, independent verification turns into reverse-engineering someone else's black box.

How to read claims like these

If you follow AI results in mathematics, it helps to keep a few simple filters in mind.

• Distinguish a preprint from a peer-reviewed publication. The first shows intent; the second confirms the result.
• Look for independent verification: outside mathematicians, formal verification, or at least a detailed walkthrough of the key steps.
• Check whether the process is disclosed. A description of how the model arrived at the proof and what humans did makes the result reproducible.
• For headline claims, such as solving a Millennium Problem, wait for official confirmation. For the Clay problems, that takes years.

For OpenAI, the stakes are reputational. If some of the 719 manuscripts survive scrutiny, that will be a strong argument for AI as a mathematician's tool. If checking turns up many errors, both these materials and trust in the company's next claims will suffer. Only the texts themselves, and the time experts spend on them, can settle the matter.