← All papers
First page of A Human Audit of OpenAIs AI-Generated Mathematical Proofs

A Human Audit of OpenAIs AI-Generated Mathematical Proofs

Mikołaj Sienicki, Krzysztof Sienicki

cs.AI Aug 3, 2026 · v3 cs.GL quant-ph
Audits OpenAI's AI-generated proofs, which come with public Lean 4 certificates, examining the repository snapshot, sorry counts, and how Lean statements correspond to the published theorems.
We assess 18 chapter-specific reviews of the ten mathematical results announced by OpenAI on 1 August 2026, alongside review standards, Lean formalizations, subsequent research, and mathematical references. The article audits this review record without claiming a complete reconstruction of all ten proofs. No confirmed substantive mathematical error in a principal result remains in the examined assessments, although review depth varies and some dependencies remain partly checked. Chapter 8 presents the strongest reservation: a specialist review requests major revision of compressed analytic arguments. In Chapter 6, an apparent polarity error was withdrawn after an overbar lost during PDF extraction was recovered from the typeset source. Subsequent research independently reuses the Chapter 3 proof mechanism and confirms that Connes's rigidity conjecture is false, without independently reproducing Chapter 4's stronger infinite-family result. Among the cited follow-ups, Chapter 7 receives the strongest direct theorem-level corroboration through a stronger hardness theorem. Related equality results in Chapter 8 do not verify the analytic inequality proof. Some follow-ups disclose material AI assistance. We argue that confidence should combine formal checking, human reconstruction, independent mathematical use, and a public record supporting correction of both proofs and reviews.

OpenAI announced ten mathematical results produced by an internal model and formalized in Lean. Favorable reviews can suggest more closure than was actually achieved. It is unclear how much of the informal and formal record was really checked.

The authors audit 18 chapter-specific reviews from a 23-document corpus. They read these alongside the AIGS review standards, the public ten-proofs Lean repository (commit 94bc0fe, Lean 4.32.0, zero sorry in the principal certificates), later research, and MathWorld entries. They separate three objects: the encoded Lean proposition, the published theorem, and the advertised claim. They also classify the kinds of evidence by what each one actually establishes.

No confirmed substantive error in a principal result survives the review record. Chapter 8 retains unresolved analytic gaps. A Chapter 6 polarity objection was withdrawn because it came from a PDF extraction error. Follow-up work supports Chapters 3, 4 and 7 to different degrees. The authors propose a cumulative confidence model: generation, formal checking, human reconstruction, independent reuse, and public correction.