OpenAI publishes 722 AI-generated math manuscripts

OpenAI released a large batch of mathematical results written by an unreleased internal model, posting 722 manuscripts grouped into 372 families on GitHub. A family collects a main result with its companion arguments and alternative proofs. The README says each result used about three hours of ChatGPT Pro thinking compute on average, and that the model was given roughly 4,000 problems over the evaluation. The topics are real research mathematics, not exercises: the irrationality exponent of pi, Kaplansky's direct-finiteness conjecture in characteristic two, the Mezard-Parisi formula for diluted spin glasses, and the isomorphism problem for free group factors, among others.

The honest part is the caveat. Not every result is formally checked. OpenAI shipped a Lean library with machine-verified formalizations for some of the work, but the README states plainly that some of the unformalized results could have issues. That matters because earlier AI math claims this year, including a disproof of an 80-year-old Erdos conjecture in May, drew both praise and sharp scrutiny from mathematicians over whether the arguments held up. Publishing the raw manuscripts and Lean files lets the community check the work directly instead of trusting a press summary.

Why it matters

If you work in mathematics or formal verification, this is a dataset you can actually audit: 722 arguments with partial Lean formalization, released for inspection rather than announced. The open question is how many survive review, and OpenAI is letting you find out for yourself.