Second mathematician accuses OpenAI of dishonesty over his ChatGPT conversations

Andreas Thom, a group theorist at TU Dresden, has accused OpenAI of "dishonesty" over how it answered his questions about whether his private ChatGPT conversations fed into its mathematical results. He says the company's reply addressed a narrower question than the one he asked, leaving the central issue unresolved. OpenAI did not respond to The Verge's request for comment on the accusation, reported on September 10.

A narrow answer to a narrower question

Thom's research overlaps with one of ten mathematical results OpenAI announced last month: work involving non-sofic groups, infinite structures that cannot be approximated by finite ones. OpenAI had already revised that writeup after criticism for failing to acknowledge contributions from Thom and fellow mathematician Gábor Kun.

He began scrutinizing his own ChatGPT interactions after NYU's Tristan Buckmaster publicly questioned whether OpenAI's models had benefited from his use of Codex. What struck Thom was how precisely the model grasped his techniques, then not the most obvious path to a solution. He wrote to Sébastien Bubeck, OpenAI's head of mathematics research, and Mark Sellke, a Harvard statistician, asking whether his conversations had been incorporated into training data or made available to its reasoning process.

The answer was confined to direct access to his conversations. It said nothing about whether they entered the training pipeline. "No such qualification, explanation, or evidence was given," Thom told The Verge. "I take this as dishonesty to say the least."

He added that "de-identification may remove a name; it does not remove the intellectual content of a mathematical idea." Of Sellke's reply he said it was "at minimum, unjustifiably broad and materially misleading; looking back it was plainly dishonest."

The second accusation in a week

The dispute lands days after a bitter public fight between OpenAI and Buckmaster over the company's claimed solution to the Navier-Stokes Millennium Prize problem, the earlier open letter and the Navier-Stokes dispute. OpenAI has maintained that its researchers and agents did not view Buckmaster's work before public release, while declining to exclude that anonymized data from his product use shaped model development; Bubeck called certain remarks "a bad choice of words."

Thom argues only OpenAI holds the records needed to answer him, so if it denies using his work the burden falls on the company to disclose the relevant datasets and clarify its data-use terms. He challenged its line between direct access and de-identified training data, noting OpenAI leaned on the same hedged framing with Buckmaster.

What would settle it

A straight answer on whether Thom's conversations entered any training run — plus disclosure of the datasets involved — would close the matter, and only OpenAI can supply either. The Clay Mathematics Institute, which administers the Millennium Prize, has said it will conduct a detailed review of the Navier-Stokes situation. Until then, this remains an accusation the company has neither confirmed nor denied.