Mathematician demands OpenAI prove it didn't train on his private ChatGPT math

Andreas Thom, a group theorist at TU Dresden, has gone public with a pointed challenge to OpenAI: prove that his private ChatGPT conversations never fed the company’s training pipeline. His concern traces back to August, when OpenAI said its Astra model had constructed the first-ever non-sofic group — a result whose central technical step leaned directly on a 2019 paper by Thom and Gábor Kun. Thom says a senior OpenAI researcher gave him a misleading answer about whether those chats touched training data, and he wants a definitive account rather than reassurances.

The sharper point Thom raises is that standard privacy tooling is the wrong tool for protecting research. Redaction and anonymization are built for personal data, so stripping a name from a transcript does nothing to protect the mathematical ideas embedded in it. A novel argument or construction remains fully usable once anonymized, which means a mathematician who workshops unpublished results with a model may effectively be handing over the work itself — the exact opposite of what ‘privacy’ controls are assumed to guarantee.

Thom is not alone. NYU’s Tristan Buckmaster and Anthropic researcher Levent Alpöge had posted documents showing advances in the same area before OpenAI’s announcement, and Buckmaster claims OpenAI became aware of the pair’s unpublished work and accelerated its own effort to reach a result first. Two mathematicians are now pressing OpenAI for answers about its training data. The episode crystallizes a growing credit-and-provenance problem in AI-assisted mathematics: when a lab claims a breakthrough that rests on others’ unpublished or barely-published work, the field has no reliable way to audit where the ideas actually came from.