OpenAI publishes 722 math proofs verified by Lean

Summary

OpenAI published on GitHub 722 AI-assisted mathematical proofs (organized into 372 theorem families) that were formally verified using Lean, addressing an open problem in Navier-Stokes, generated by the same model in July 2026.

Why it matters

Formal verification with Lean eliminates the 'plausible but wrong' error mode that frequently affects large language model (LLM) outputs. This ensures that proofs are not only persuasive but objectively machine-checkable. This marks a shift of AI from a research assistant to a verifiable scientific partner. The community now faces questions regarding authorship, citation norms, and the future role of human mathematicians as sole arbiters.

Highlights

• 722 proofs covering active research areas such as analytic number theory, commutative algebra, differential topology, and algebraic geometry. These are not textbook exercises.
• Each proof consumed approximately 3 hours of extended reasoning compute. The entire batch required over 2,100 hours of processing time.
• Buckmaster from Princeton University and the AGMAI advisory panel warn that while Lean confirms logical consistency, it does not judge mathematical depth or whether the original open question is truly addressed.

Critical take

The article's triumphalist tone slightly exaggerates the achievement. Lean can verify that a proof contains no gaps, but it cannot judge whether the theorem is new, interesting, or actually addresses the original open problem. This distinction, acknowledged by the author, is still presented in celebratory language. Furthermore, the Navier-Stokes result underpinning the 'consistent model' narrative is still undergoing peer review. Therefore, the broad claim of reliability relies partially on an unverified assumption.