OpenAI has released 719 mathematics manuscripts containing nearly 400 claimed AI-generated results, creating a review task that researchers say could take years. The collection spans geometry, number theory, algebra, topology, probability, theoretical computer science and mathematical physics.
The volume is not the same as verification. OpenAI acknowledges that the manuscripts are at different stages, and only about 300—roughly 42 percent—currently have formalizations in Lean. Lean is a proof assistant that can mechanically check whether a stated conclusion follows from encoded assumptions.
Even a successful Lean check does not automatically establish that a formal statement captures the important claim in the accompanying paper. Mathematicians must inspect definitions, assumptions, novelty and links to existing literature. Researchers who spoke with The Verge also reported uneven writing and formalization quality across the repository.
The release may contain important new work, but its immediate effect is to shift a large validation cost onto the mathematics community. OpenAI says it will add more formalizations as they become available. Until specialists review individual results, the collection should be treated as a mixed set of candidate contributions rather than hundreds of equally established discoveries.