OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
A sign error in one algebraic geometry proof forced OpenAI to withdraw papers from its 722-manuscript math release, reviving questions about verification
OpenAI released 722 AI-generated mathematics papers on October 6, 2026. By October 7, three of them were gone.
The company withdrew the papers after finding a sign error in one proof. That single mistake undercut key arguments in two related manuscripts that depended on it.
What OpenAI released, and what it pulled #
The original drop landed on OpenAI’s GitHub repository. It contained 722 papers grouped into 372 result families, all produced by an unnamed proprietary model.
The manuscripts aimed at around 4,000 unresolved problems in mathematics. They claimed advances on numerous longstanding conjectures, including progress on several Millennium Prize Problems.
The model reportedly spent approximately three hours on average per result.
The flawed proof sat in algebraic geometry. A sign error means a plus where a minus should be, or the reverse. In a proof, one flipped sign can turn a valid chain of logic into one that proves nothing at all, and every argument built on top of it falls with it.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
After the withdrawals and revisions, the corrected catalog now stands at 719 papers. OpenAI emphasized that the mathematical assertions themselves have not been refuted.
The verification problem #
After the initial release, 300 of the 719 results had been formally verified in Lean, a proof assistant that translates a proof into formal code and checks it step by step. That works out to about 42%.
Critics also pointed to the proprietary nature of the model. Outside researchers cannot inspect it or reproduce its process.
The release was informed by the Advisory Group on Mathematics and Artificial Intelligence, known as AGMAI.
A pattern of friction with mathematicians #
This is not OpenAI’s first clash with the math community. The company previously withdrew a $1M sponsorship from a Caltech AI math contest after criticism from 771 mathematicians.
What this means #
The formally verified results carry a strong guarantee. The remaining majority carry the same kind of risk that produced the sign error.
OpenAI’s insistence that the claims remain unrefuted is technically meaningful but limited. An unrefuted claim with a broken proof is, mathematically speaking, back to being a conjecture.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our