A mathematician says OpenAI's answer about his private chats was evasive

Anonymization strips a name, not an idea. That distinction sits at the center of a dispute between a group theorist and OpenAI.

The sequence

On August 1, 2026, OpenAI announced that its model had constructed the first known non-sofic group, settling a question open since Mikhail Gromov posed it 27 years earlier. The central technical step of that proof leans on a 2019 paper by Gabor Kun and Andreas Thom.

Thom, who works at TU Dresden, had spent months with a Dresden colleague working through the same problem inside ChatGPT. Shortly after the announcement he emailed OpenAI researchers Mark Sellke and Sebastien Bubeck with two questions: whether those exchanges entered training data, and whether the system could reach them while working on the proof.

Sellke’s reply, as Thom quotes it, was a single line: “that did not happen.”

What Thom is actually claiming

Thom does not assert that OpenAI trained on his conversations. His argument is narrower and harder to dismiss. The reply addresses only the second question, direct access during the solving process, while leaving the first one untouched. He describes it as unjustifiably broad and, in hindsight, dishonest, and has asked the company to state plainly that the training data claim is also false.

OpenAI has not publicly responded to his posts.

The norm that does not exist yet

A researcher who uses these tools to think is potentially routing unpublished work through a system that may later reach the same result. Academia has centuries of established practice around priority and credit, and none of it was written for a case where the competitor is also the notebook.