What happened
On 6 October 2026, OpenAI put 722 maths papers online. Its AI had written them, and OpenAI said they tackled 372 open maths problems.
The next day, OpenAI checked its own work and took three of the papers back. One proof used +1 where, by its own rules, it should have used -1. That one small slip broke the proof. Two other papers were built on top of it, so they fell too.
OpenAI also fixed 14 more papers, with “proof repairs, corrected statements, clearer hypotheses and dependencies.”
Why it went wrong
Maths proofs are long chains of steps. If one step is wrong, everything after it is wrong too. A wrong sign is easy to miss, for a person or for a machine. Here, the AI made the slip, and the papers went public before anyone caught it.
Impact
- Three papers withdrawn and 14 corrected within a day.
- No harm outside the maths world, but it raised doubts about checking 722 papers this quickly.
The lesson
- More is not better if nobody has checked it. Hundreds of papers at once is too many to check properly.
- Small slips can spread. One error took down three papers.
- Checking your own work counts. OpenAI found this itself and said so in public, which is the right way to handle it.