Andreas Thom Accuses OpenAI of Dishonesty Over 10 Math Results

3 articles · Updated · The Verge · Sep 10

Andreas Thom said OpenAI gave a “materially misleading” answer when he asked whether his ChatGPT interactions could have fed into models behind a non-sofic groups result announced last month.
One of the 10 touted breakthroughs drew heavily on prior work by Thom and Gábor Kun, and Thom said OpenAI’s unusually detailed grasp of their techniques deepened his suspicion.
OpenAI had already quietly amended its writeup after criticism that it failed to credit recent contributions, but Thom says the company still has not proved user interactions were excluded from training data.
The dispute echoes OpenAI’s stance in a separate Millennium Prize controversy, where it denied using specific user data yet would not rule out de-identified usage improving its models.
That ambiguity is fueling wider unease among mathematicians, who fear unpublished ideas and even rumors of progress could trigger races with AI firms and make the field more secretive.