Andreas Thom, a mathematician whose work on non-sofic groups became one of OpenAI's ten announced breakthroughs, is demanding the company prove it did not use his unpublished research to train its models. His accusation, posted on Mastodon in recent days, follows a similar complaint from Tristan Buckmaster, a mathematics professor at New York University, and adds to a growing pattern of researchers questioning where OpenAI gets the data behind its mathematical advances.
The core of the dispute is simple. Thom had interactions with ChatGPT that touched on his research area. OpenAI then announced a result in that exact area, initially failed to credit his prior work, and later amended its writeup after criticism. When Thom emailed OpenAI researchers Sébastien Bubeck and Mark Sellke to ask whether his conversations had been part of the training data or accessible to the reasoning process, the answer sidestepped the question he actually asked.
The distinction that matters
OpenAI told Thom it could confirm that his specific conversations with ChatGPT had not been accessed directly to produce the non-sofic groups result. What it would not say was whether de-identified data derived from those conversations had entered the training pipeline that improved the underlying models. This is the same distinction the company drew when it announced its solution to the Navier-Stokes Millennium Prize problem, one of mathematics' most famous unsolved questions.
In that case, OpenAI researchers and agents said they had not seen any work by Buckmaster and Anthropic researcher Levent Alpöge, who were working on the problem in a personal capacity, until it was released publicly. But the company added a qualification: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." Thom called this the same obfuscatory move. De-identification strips a name from a conversation. It does not strip the intellectual content of a mathematical idea from the text.
Non-sofic groups, for context, are infinite mathematical structures that cannot be approximated by finite ones. They sit in a specialized corner of group theory. OpenAI's result in this area was one of ten it announced with considerable fanfare. The company acknowledged that its solution built heavily on previous work by Thom and fellow mathematician Gábor Kun, but the original writeup did not credit them. After widespread criticism in mathematical circles, OpenAI quietly amended its announcement.
Why researchers cannot check for themselves
Thom pointed out that mathematicians are not equipped to reverse-engineer OpenAI's training pipeline. They cannot audit the datasets, trace the flow of data through the model, or determine whether specific interactions contributed to training. Only OpenAI has the relevant information. If the company is going to deny using a researcher's work, Thom argued, the burden of proof is on OpenAI to disclose the necessary datasets and clarify the terms under which it uses data.
Mark Sellke, a statistician at Harvard who is also an OpenAI researcher, gave Thom what he described as a "categorical answer." Looking back, Thom called that response "at minimum, unjustifiably broad and materially misleading; looking back it was plainly dishonest." The company did not immediately respond to requests for comment on Thom's accusations.
Buckmaster's complaint, which came first, centered on his use of OpenAI's Codex. He raised questions about whether the company's AI models had benefited from his work. The pattern across both cases is the same: a researcher interacts with an OpenAI product, OpenAI publishes a breakthrough in the researcher's area, and the company declines to rule out indirect influence while denying direct access to user data.
The broader chilling effect
Multiple researchers told The Verge that incidents like these worry them. If mathematicians know that even rumors of progress on a major problem could trigger a race with a well-resourced technology company, the field may become more secretive. Researchers may stop sharing preliminary results, delay publications, or avoid discussing work with anyone connected to companies that could turn that information into competing research.
This concern is not hypothetical. OpenAI said it pursued the Navier-Stokes problem in the first place because it heard rumors online that other researchers had made significant progress. The company decided to try as well. For mathematicians who have spent years or decades on a problem, learning that a technology company entered the race based on hearsay and potentially benefited from their prior interactions with AI tools is deeply unsettling.
The mathematics community has already seen what happens when trust breaks down. The initial failure to credit Thom and Kun in the non-sofic groups announcement required public pressure to fix. The subsequent refusal to conclusively rule out training data influence suggests OpenAI views the question of data provenance as something to manage rather than resolve.
For developers and researchers who use AI tools in their work, Thom's argument raises practical questions. If you use ChatGPT, Codex, or any similar product to discuss your research, brainstorm approaches, or debug code, the data you provide may flow into training pipelines that improve future models. Those models may then be used by the same company to compete with your work. The de-identification process removes identifiers but preserves the substance of what you discussed.
Thom called it ethically indefensible if nonpublic research supplied by users helped improve models that the company then used to race those same users to publication, without consent, proper disclosure, or credit. The question now is whether OpenAI will address the substance of that claim or continue drawing distinctions that, as Thom put it, separate the name from the idea while leaving the idea itself up for grabs.