OpenAI Withdraws Three AI-Generated Mathematics Papers After Discovering Proof Errors

OpenAI has withdrawn three research manuscripts from a major collection of artificial intelligence-generated mathematical findings after identifying an error that invalidated part of the underlying mathematical argument. The company published more than 700 manuscripts on October 6, 2026, presenting results produced by an internal frontier AI model across several areas of mathematics. However, on October 7, the company announced that three papers had been removed following the discovery of a sign error in one proof. The mistake affected an argument involving the algebraicity of certain mathematical structures and also undermined constructions used in two dependent papers. The withdrawals demonstrate the importance of independently checking AI-generated mathematical claims before treating them as established scientific results.

The three withdrawn manuscripts were part of a larger research release covering problems in algebra, geometry, number theory, theoretical computer science and other mathematical disciplines. One paper examined the algebraicity of Weil classes on split abelian eightfolds, while two related manuscripts addressed mathematical results involving Kuga–Satake correspondences for K3 surfaces and the rational Hodge conjecture for products of K3 surfaces. According to the repository’s revision notes, the sign error invalidated a key cancellation argument and affected the constructions used in the related studies. Because the papers depended on the flawed reasoning, correcting the original manuscript alone was not sufficient. OpenAI therefore withdrew all three rather than allowing the affected results to remain presented as valid findings.

The company also revised 14 other manuscripts to address proof problems, clarify mathematical statements, strengthen assumptions and correct references. OpenAI said it identified the errors during an audit and intends to continue updating the research repository as additional corrections or formal proofs become available. The company has encouraged mathematicians to examine the results and provide feedback, describing the release as part of an effort to accelerate mathematical discovery. Nevertheless, the episode illustrates a central challenge in using AI for advanced research: a model may produce an argument that appears sophisticated and technically detailed while still containing a small error capable of invalidating the conclusion. In complex mathematics, a single incorrect sign, assumption or logical step can undermine an entire proof or several dependent results.

Another major issue is the extent to which the mathematical findings have been formally verified. OpenAI included formalizations of many proofs using Lean, a programming language and proof assistant that enables computers to check whether mathematical arguments satisfy specified logical rules. However, the company acknowledged that formalizations were not available for every result at the time of publication. Reporting by Retraction Watch indicated that roughly half of the released results had not been formally confirmed when they were first published. Mathematicians have consequently urged readers to distinguish between promising AI-generated arguments and results whose correctness has been rigorously established. Formal verification can provide strong evidence that a proof is logically sound, but the formalized statement must also accurately correspond to the mathematical claim made in the accompanying paper.

The withdrawals come amid a wider debate about how AI companies should publish major scientific findings. Supporters argue that advanced AI models could help mathematicians explore difficult problems, discover new approaches and accelerate research that might otherwise take years. Critics, however, warn that releasing hundreds of technical manuscripts before their arguments have been fully checked can place an unreasonable burden on researchers and create confusion about which results are reliable. OpenAI has said it will improve the clarity, citations and presentation of future publications while continuing to work with an independent advisory group on mathematics and artificial intelligence. The latest corrections do not establish that the entire collection is unreliable, but they reinforce the need for transparent documentation, independent scrutiny and rigorous verification. As AI systems become more involved in scientific discovery, the credibility of their contributions will depend not only on the number of results they produce, but also on whether those results withstand careful mathematical examination.