Andreas Thom, a mathematician at TU Dresden, accused OpenAI of plagiarizing his 20-year research on non-sofic groups in the new Astra model. Thom claims the model's proof of the 27-year-old Gromov's soficity conjecture matches his unpublished work.
The researcher and his collaborator Gábor Kun had previously discussed their unpublished findings and proof details with ChatGPT over several months. Thom sent an inquiry to OpenAI researchers Mark Sellke and Sébastien Bubeck regarding the use of these logs in model training.
Mark Sellke denied the allegations in a brief email. Thom criticized this response as dismissive, noting that it failed to clarify whether his data was used for training or retrieved during the model's inference process.
OpenAI later acknowledged in a separate dispute involving mathematicians Tristan Buckmaster and Levent Alpöge that it uses de-identified data from its products to improve models. Thom argues that de-identification does not protect the intellectual core of mathematical ideas.
This incident follows a similar conflict between OpenAI and Tristan Buckmaster regarding the Navier-Stokes equations. Critics, including Valerio Capraro, have characterized these recurring accusations as a significant intellectual property controversy.
