Mathematician Andreas Thom has publicly accused OpenAI of lacking transparency regarding the data used to train its models. Thom raised concerns that his private interactions with ChatGPT, which involved his research on non-sofic groups, may have influenced the company's recent mathematical results.
The controversy began after OpenAI announced a solution involving non-sofic groups, a field where Thom and Gábor Kun had previously conducted research. Although OpenAI later updated its documentation to acknowledge their work, Thom remains critical of the company's data practices.
Thom reached out to OpenAI researchers Sébastien Bubeck and Mark Sellke to inquire if his inputs were used in training. He found their responses unsatisfactory, noting that they failed to clarify whether his intellectual content was ingested into the company's broader training datasets.
This dispute follows similar concerns raised by New York University professor Tristan Buckmaster, who questioned if his use of OpenAI's Codex contributed to the company's AI models. OpenAI has stated it did not access specific user data for its breakthroughs but could not rule out indirect influence from de-identified data.
Thom argues that de-identification does not protect the intellectual content of mathematical ideas. He suggests that OpenAI should disclose its datasets to prove that user research was not used to gain a competitive advantage in publishing mathematical discoveries.
