Mathematicians want proof OpenAI didn't use their work
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Mathematicians are demanding OpenAI prove its math breakthrough didn't use their unpublished work.
The fight is over who gets credit when AI solves a famous problem.
Background
OpenAI announced last month that its models had solved the Navier-Stokes Millennium Prize problem, an achievement it said involved 2.7 million messages and roughly 130 billion output tokens. Mathematicians Tristan Buckmaster and Levent Alpöge had used OpenAI's Codex as paying customers to find an almost identical solution to one part of the problem, and their work was stored on OpenAI's servers. When OpenAI's announcement came, Buckmaster and others questioned whether their unpublished work had fed into the model, and OpenAI denied it. The dispute has since widened, with a second mathematician, Andreas Thom, raising similar transparency concerns.
How it unfolded
- recentlyMathematicians Tristan Buckmaster and Levent Alpöge use OpenAI's Codex as customers to discover an almost identical solution to one portion of the Navier-Stokes problem, without a full proof.
- Sep 8, 2026OpenAI announces its models solved the Navier-Stokes Millennium problem, saying the effort used 2.7 million messages and about 130 billion output tokens, and would have cost a customer $15 million. At a news conference, OpenAI's Sébastien Bubeck denies using Buckmaster and Alpöge's prompt or proof.
- Sep 9, 2026Buckmaster posts a document saying he is not accusing anyone and has not seen OpenAI's proof, but calls for an independent investigation into how OpenAI's work began. OpenAI's Mark Chen says he is 'a little disappointed with the allegations.'
- Sep 10, 2026Mathematician Andreas Thom raises concerns on Mastodon that his and colleagues' ChatGPT interactions may have contributed to OpenAI's result, and says only OpenAI has the data to prove otherwise.
Who’s saying what
- OpenAI
- Sébastien Bubeck denies the allegations as 'false and inflammatory,' saying the team did not use Buckmaster and Alpöge's prompt or proofs and did not see their work until it was released publicly.
- Caution
- Buckmaster says he is not accusing anyone and has not seen OpenAI's proof, but calls for an independent investigation into how OpenAI's work began.
- Public
- Andreas Thom says researchers cannot reverse-engineer OpenAI's training pipeline and that only OpenAI has the relevant data, so the company should disclose datasets and terms if it denies using their work.
Still unverified
Whether OpenAI's models trained on or accessed Buckmaster and Alpöge's work remains unconfirmed; OpenAI says it cannot rule out that de-identified data from their usage helped improve its models, and Buckmaster says he does not know whether their data was used. OpenAI's paper cites 29 sources, but there is no way to know if that list is complete.