OpenAI’s October 6 release of AI-generated mathematical manuscripts has left researchers with a new task: deciding which claims they can verify and how to understand the rest. At an October 8 seminar organized by the French Mathematical Society, Fields Medalist Hugo Duminil-Copin discussed that challenge.
Le Monde reported on October 12 that 16 of the open questions OpenAI’s model tackled were related to Duminil-Copin’s work. In a separate point, the newspaper said he plans to examine more than 3,500 pages of AI-generated material related to his area and organize collective work around the results. It also reported his concern that competition to claim AI-generated solutions could discourage mathematicians from openly discussing research questions.
Checking whether a result holds is different from producing a manuscript that presents a proof. OpenAI said its release contained 722 manuscripts across 372 families of related results. Its repository lists different verification stages and says many, but not all, manuscripts have Lean formalizations, which let software check specified proof steps. The repository warns that some unformalized results may contain errors. Formalization can help, but mathematicians still need to understand and verify the arguments.
Sources
- Le Monde: report on Duminil-Copin’s response, October 12, 2026
- OpenAI: manuscript release description, October 6, 2026
- OpenAI’s manuscript and proof-artifact repository
Image: MathKnight, mathematics blackboard at Tel Aviv University, CC BY 2.5. Illustrative image; not a photograph of the seminar.