Andreas Thom accuses OpenAI of "dishonesty" over math data

This is the second data-transparency fight in OpenAI's math push in a matter of days. The first centered on Tristan Buckmaster, a mathematics professor at New York University, who publicly questioned whether the company's models had benefited from his use of OpenAI's Codex while he worked on problems with Anthropic researcher Levent Alpöge. Thom said he began reflecting on his own ChatGPT interactions only after that challenge went public, which is how one dispute fed directly into the next.

The non-sofic groups episode is the concrete link. After OpenAI announced the result with fanfare, mathematicians criticized it for failing to acknowledge recent contributions from Thom and Kun, and the company quietly amended its writeup. OpenAI's own blog post on its Navier-Stokes solution drew the same fine line: it denied that the researchers and agents saw any specific user work before public release, yet would not rule out that de-identified data from users' product usage helped improve its models. Thom's reply is that de-identification removes a name but not the intellectual content of a mathematical idea.

What gives the dispute its weight is the asymmetry Thom describes: only OpenAI holds the data needed to settle the question, so the burden of proof sits with the company rather than the researchers. The stakes for mathematicians, per researchers who spoke to The Verge, are that rumor alone of a near-breakthrough could ignite a race with a well-resourced tech giant, pushing the field toward secrecy. OpenAI did not immediately respond to The Verge's request for comment, so whether it discloses datasets and data-use terms is likely to shape how openly mathematicians share ideas going forward.