Following the controversy surrounding the creation of OpenAI’s Navier-Stokes proof, another mathematician is now accusing the company. Andreas Thom, Professor of Geometry at TU Dresden, suspects that unpublished research content from his ChatGPT conversations may have been incorporated into the training of OpenAI models. This could potentially have contributed to another mathematical breakthrough that OpenAI presented in early August.
According to the company, an internal version of Astra had, for the first time, answered a central question in group theory. The proof significantly builds on earlier work by mathematician Gábor Kun from the Alfréd-Rényi Institute in Budapest, as well as on joint work by Kun and Thom. Thom found it suspicious that Astra pursued this particular line of research, even though, according to his account, it was not considered the most obvious solution path.
Thom demands transparency on data usage
As Thom writes on Mastodon, he had previously spent months actively discussing ongoing research questions with a colleague in Dresden using ChatGPT. Following OpenAI’s August announcement, Thom asked OpenAI researchers Mark Sellke and Sébastien Bubeck two questions: Firstly, whether content from these conversations had been included in training data, and secondly, whether the conversations had been accessible to the AI system during its search for the proof. According to Thom, Sellke merely replied, “Regarding your conversations with ChatGPT: that did not happen.”
