Mathematician Accuses OpenAI of Using Unpublished Research in Model Training
Mathematician Andreas Thom accused OpenAI of incorporating unpublished research from his private ChatGPT conversations into model training without consent. Thom said OpenAI's response to his inquiries was misleading and demanded the company disclose details of its data use practices.
What happened
Andreas Thom, a mathematician at TU Dresden, accused OpenAI of incorporating unpublished research from his private ChatGPT conversations into its model training without consent. Thom said he had spent months discussing ongoing research questions with colleagues using ChatGPT before OpenAI announced a breakthrough in group theory in August 2026.
OpenAI's Astra system solved a problem in non-sofic groups that built heavily on work by Thom and Gábor Kun. Thom found it suspicious that Astra pursued this particular research path, which he said was not considered the most obvious solution at the time, suggesting knowledge of his unpublished discussions.
Thom wrote to OpenAI researchers Sébastien Bubeck and Mark Sellke asking whether his ChatGPT conversations had been part of the training data or accessible to the reasoning process. Sellke replied that Thom's conversations with ChatGPT "did not happen," which Thom later characterized as "unjustifiably broad and materially misleading."
Reported impact
Thom accused OpenAI of unethical and "dishonest" behavior and lack of transparency about its training data origins. He demanded OpenAI disclose which data, model versions, and account settings its denial was based on if the company continued to deny his claims broadly.
Index analysis
This is the second accusation against OpenAI over mathematical training data, following a controversy involving mathematician Tristan Buckmaster regarding use of his Codex interactions. OpenAI previously distinguished between two cases: researchers did not access specific user data from mathematicians, but de-identified data from their previous use of ChatGPT may have contributed to model improvement.
Governance controls
Thom deactivated use of his content for model training on June 29, but noted this left unclear what happened with older conversations or previously derived data. Thom stated it was not possible to verify from outside how such a setting is technically implemented.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.