Second Mathematician Accuses OpenAI of Dishonesty Over Training Data

Mathematician Andreas Thom has publicly accused OpenAI of failing to provide a clear answer about whether his private interactions with ChatGPT could have contributed to one of the company’s recent mathematical breakthroughs. The allegations add to growing scrutiny surrounding OpenAI’s claims of major advances in mathematical research and the data used to develop its AI systems.

Thom’s concerns focus on OpenAI’s work involving non-sofic groups, an area of mathematics in which he and fellow researcher Gábor Kun have made significant contributions. OpenAI previously acknowledged that its result relied heavily on their earlier work, while also facing criticism for initially failing to properly acknowledge those contributions before revising its public write-up.

According to Thom, he contacted OpenAI researchers Sébastien Bubeck and Mark Sellke after reflecting on conversations he and his colleagues had held with ChatGPT. He wanted to know whether those conversations had entered OpenAI’s training data or were otherwise accessible to the system responsible for the mathematical result. Thom said the response addressed whether the conversations had been directly accessed during the research process but did not fully answer whether the material could have been incorporated into training data.

The dispute comes shortly after a separate controversy involving mathematician Tristan Buckmaster, who questioned whether information from his unpublished research may have influenced OpenAI’s claimed work on the Navier-Stokes problem. OpenAI has denied directly accessing specific user work before it became public, while acknowledging that it could not completely rule out the possibility that de-identified data from product use had contributed to model improvement.

The allegations have raised broader questions about data provenance, research ethics and intellectual credit as AI systems become capable of producing increasingly sophisticated mathematical work. Thom argues that researchers cannot independently determine whether their private interactions were incorporated into training systems because the relevant information is held by OpenAI. The company had not publicly responded to The Verge’s request for comment at the time of its report, leaving the underlying question unresolved.