Second Mathematician Accuses OpenAI Over Its Training Data

A second mathematician has accused OpenAI of stonewalling him on whether his private work with ChatGPT fed into one of the company's celebrated math results. Andreas Thom, a mathematician at TU Dresden, published a series of Mastodon posts Wednesday laying out email exchanges with OpenAI researchers Sébastien Bubeck and Mark Sellke, and called the company's answers "dishonest."

The complaint lands two days after a similar, separate dispute broke over OpenAI's headline-grabbing Navier-Stokes result, and it is not a rehash of that story. Thom's accusation involves a different mathematician, a different result (non-sofic groups, not fluid dynamics), and a specific, documented back-and-forth over email that he has now made public.

RelatedMathematician Says OpenAI Pressured Him Over Navier-Stokes Credit

What did Andreas Thom actually allege?

Last month OpenAI announced ten new mathematical results produced with help from its models. One of them concerned non-sofic groups, infinite algebraic structures that can't be approximated by finite ones, and it built heavily on prior work by Thom and fellow mathematician Gábor Kun. Mathematicians in the field criticized OpenAI for failing to credit that work properly, and the company quietly revised its writeup after the pushback.

That's what got Thom looking back at his own history with the chatbot. He said he was struck by "OpenAI's detailed command of our techniques," ones he described as neither obvious nor the most promising path to the result at the time. So he emailed Bubeck and Sellke directly and asked a narrow, specific question: had his conversations with ChatGPT become part of the training data, or been otherwise accessible to the model's reasoning process, in a way that could have contributed to OpenAI's result?

The reply he got, he says, answered a question he hadn't asked. It addressed whether his chats could be directly retrieved, not whether they had entered the pool of data OpenAI uses to train and improve its models. "No such qualification, explanation, or evidence was given," Thom wrote. "I take this as dishonesty to say the least."

How does this connect to the Navier-Stokes dispute?

Thom's posts didn't come out of nowhere. Days earlier, NYU mathematician Tristan Buckmaster went public questioning whether OpenAI's models had drawn on his use of the company's Codex tool. Buckmaster and Anthropic researcher Levent Alpöge had been quietly working, in a personal capacity, on problems related to the Navier-Stokes equations, the fluid-dynamics problem that carries a $1 million Clay Millennium Prize. OpenAI then announced its own Navier-Stokes solution and, in its blog post, wrote flatly that "we did not see any of their work through any means until they released it publicly."

That denial came with a carve-out that Thom later recognized in his own exchange with the company: OpenAI said it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." Thom calls that the same evasive move Sellke made with him. "De-identification may remove a name," he said. "It does not remove the intellectual content of a mathematical idea."

1. Aug 2026OpenAI announces ten AI-assisted math results, including a non-sofic groups proof. Mathematicians flag missing credit to Thom and Kun; OpenAI quietly amends the writeup.
2. Sep 8, 2026Tristan Buckmaster goes public over the Navier-Stokes credit dispute. OpenAI denies direct access to his unpublished work, hedges on indirect use.
3. Sep 10, 2026Andreas Thom publishes his own email exchange with OpenAI on Mastodon. Calls the company's answers "dishonest" and demands disclosure.

Why won't OpenAI give a straight answer?

Because a straight answer is hard to give without exposing exactly what the training data contains, and OpenAI has never published that list for any of its recent models. Thom's point, and it's a sharp one, is that researchers have no way to check the claim themselves. "Only OpenAI has the relevant data for that," he wrote. If the company wants to deny using nonpublic research, the burden falls on it to prove it, not on outsiders to disprove a negative.

RelatedClaude Formalizes First Full Proof of Fermat's Last Theorem

OpenAI did not immediately respond to a request for comment on Thom's allegations.

What does this mean for researchers using AI tools?

The practical fear among mathematicians The Verge spoke with isn't really about credit lines. It's that the incentive now runs backward. If sharing a hard, unsolved problem with a chatbot might later let a well-funded lab race you to publication using your own ideas, the rational move is to stop sharing early work at all, even informally, even in a private chat. Several researchers said exactly that: this episode could push mathematics into a more secretive, closed posture at a moment when open collaboration is supposed to be the field's strength.

There's a commercial angle too. OpenAI is actively selling its math and reasoning capabilities to research labs, universities and enterprise customers as a productivity multiplier. A running narrative that the company can't or won't confirm what's in its training data undercuts that pitch with exactly the technical, detail-oriented customers it needs to trust it most.

Our take

OpenAI's actual mathematical results here look real and hard-won; nobody serious is disputing that the non-sofic groups proof or the Navier-Stokes work required genuine capability. The problem is narrower and more damaging than that: a company that built its research reputation on rigor is answering a yes-or-no data question with language engineered to survive a fact-check rather than resolve one. "We cannot rule out" is not a denial, and mathematicians, of all people, are trained to notice the difference.

Original analysis by GenZTech Team.