Math Advisers Say OpenAI's Flood of Proofs Broke the Field's Rules

TechCrunch reports that a wave of mathematical proofs released by OpenAI did not meet the rules drawn up by a group of research mathematicians. The lab itself consulted that group. The report contains no numbers: neither the number of proofs nor the number of papers with objections. Still, the picture is clear. A company that sells its models as a tool for science has received a public rebuke from a community whose recognition it needs.

What is known so far

The source offers a single statement: OpenAI's "stream" of proofs departed from the guidelines set by advisers from the mathematical community. Everything else has to be pieced together from context. I will not fill in details that the article does not contain.

• Date and source: TechCrunch, October 8, 2026.
• Subject of the complaint: the mass release of proofs prepared by OpenAI.
• Substance of the complaint: non-compliance with rules agreed by the advisory mathematicians.
• Not disclosed: the number of papers, the specific rules at issue, the list of disputed proofs.
• Not disclosed: OpenAI's official response.

Why mathematicians are so particular

In mathematics, a result is worth exactly as much as its verification. A proof must be complete and reproducible, with clear attribution of authorship and grounding in the known literature. If proofs arrive in a batch without an agreed procedure, reviewers have to dig through the pile by hand. And the experts who can evaluate a complex proof number only in the dozens on any narrow topic. Their time is limited, and publication guidelines exist precisely to save it.

So the phrase "departed from the guidelines" is harsher than it sounds. The dispute is not necessarily about whether an individual theorem is correct. It is about whether the channel through which results reach people can be trusted at all.

Not the first episode

OpenAI has already had public disputes with mathematicians. In October 2025, company employees said GPT-5 had "found solutions" to a number of Paul Erdős's problems. Mathematician Thomas Bloom, who maintains the catalog of those problems, objected: the model had located existing publications and had not produced new solutions. The post was deleted, and the industry debated how announcement wording runs ahead of verification. In the summer of 2025, the gold medal result at the International Mathematical Olympiad also drew dispute, including over the timing of the announcement.

Against that background, the new story looks like a continuation of a pattern. Language models are indeed becoming more useful in formal reasoning, and mathematicians themselves acknowledge it. But at the lab, the speed of product releases regularly diverges from the pace of verification that science sets.

How to read such claims

For a reader, an investor, or a developer building a product on models' mathematical capabilities, there are a few practical takeaways.

• Look at who did the checking. Independent expert review or formal verification, for example in Lean, carries more weight than a press release.
• Look for the procedure. Is there a public protocol: which problems, which constraints, how the contributions of the model and of humans are recorded.
• Separate the types of results: a new proof, a restatement of a known one, a literature search. These are different achievements.
• Wait for OpenAI's response and the list of disputed papers. It is not yet known whether this involves isolated formatting lapses or a systemic problem.

If the lab wants its models to be regarded as a scientific tool, it will have to live by the rules of science, including those written by its own advisers. Otherwise, every next announcement will be met not with the question "what's new?" but with "what's wrong this time?"