OpenAI Withdraws Three AI-Generated Math Papers Following Sign Error

OpenAI released 722 AI-generated mathematics papers on October 6, 2026, but by October 7, three of them were withdrawn. The retraction was prompted by the discovery of a sign error in one proof, which undermined key arguments in two related manuscripts.

The original release, hosted on OpenAI’s GitHub repository, included papers grouped into 372 result families, addressing around 4,000 unresolved problems in mathematics and claiming advancements on several longstanding conjectures, including some Millennium Prize Problems. Each result reportedly took about three hours of compute time.

The flawed proof originated from algebraic geometry, where a single sign error can invalidate an entire logical argument. Following the withdrawals, the corrected catalog now contains 719 papers. OpenAI clarified that the mathematical assertions themselves have not been refuted.

Out of the remaining papers, 300 have been formally verified using Lean, a proof assistant that checks proofs step by step, representing about 42% of the total. Critics have raised concerns about the proprietary nature of the model used to generate these papers, as it limits external researchers' ability to inspect or reproduce the results.

This incident marks another point of friction between OpenAI and the mathematics community, following previous controversies, including the withdrawal of a $1 million sponsorship from a Caltech AI math contest after criticism from mathematicians.