Dresden Mathematician Accuses OpenAI's Astra of Plagiarizing Unpublished Non-Sofic Group Work

In early August 2026, OpenAI announced that its next-generation model Astra had solved ten mathematical problems it described as Fields Medal-level, including the Gromov flexibility conjecture, which had troubled group theory for 27 years. Astra said it had found evidence for the existence of non-flexible groups. Andreas Thom, a mathematician at TU Dresden, and his long-time collaborator Gabor Kun subsequently found that the most central key derivation step in Astra's proof closely matched the technical path the two were pursuing.

Thom disclosed that in the months before OpenAI formally released its results, he and his colleague had been using ChatGPT intensively to discuss the extended matching problem, and had entered a large amount of unpublished core derivations and the latest proof details concerning the conjecture into the chat window. Thom turned off the model training option on June 29, 2026, but he believes that turning off the switch can only apply to the period afterward, and cannot explain what happened to earlier conversations, nor whether derivative data had already been selected into the training process.

After noticing the anomaly, Thom sent a formal email to two core OpenAI researchers, Mark Sellke and Sebastian Bubeck, raising two questions: whether the conversation records discussing unpublished results in ChatGPT over several months had been incorporated into the model's training data; and whether, during the model's problem-solving and reasoning process, these records could be directly retrieved and accessed. Sellke replied: "Regarding your conversations with ChatGPT: that did not happen." Thom believes that this response provided no data audit explanation or explanation of privacy mechanisms.

Previously, New York University mathematician Tristan Buckmaster clashed with OpenAI over attribution for the Navier-Stokes equations. In response to questions from Buckmaster and Levent Alperg, OpenAI's official position subsequently changed: while insisting that it had not accessed the private data of specific accounts, it acknowledged that it "cannot rule out the use of de-identified data from customer products to improve the model."