What happened
On October 6, 2026, OpenAI published on the GitHub repository openai/math mathematical results from an unreleased internal model: 372 result families and 722 manuscripts (719 after withdrawals). It says the model was given roughly 4,000 open problems and each kept result cost about three hours of ChatGPT Pro thinking on average. About 42% of the top-line results are machine-checked in the proof language Lean; the rest were released without human or machine confirmation.
The next day, October 7, a single sign error invalidated three manuscripts (algebraic Weil classes, Kuga-Satake correspondences for K3 surfaces, and the rational Hodge conjecture for products of K3 surfaces), which were withdrawn, and 14 more were revised. The IAS advisory group had recommended releasing the model, the prompt and per-problem compute — none of which happened. The Association for Human Mathematics urged mathematicians to stop collaborating with OpenAI.
Why it matters
The speed of producing results has outrun the speed of reading and verifying them. As Scott Aaronson noted in commentary covered by outlets, almost none of the proofs had yet been understood by a human. In engineering terms, it is like merging 700 PRs without review.
What to do now
- In any AI-output pipeline, set a verification gate (human, Lean/tests, or both) before treating a result as fact.
- Prefer announcing first what is already machine-checked.
- Treat fast retractions as healthy — and community pressure as a reputation warning.