Researchers Suspect OpenAI and Anthropic of Stealing Their Discoveries

A growing group of researchers thinks their private chats with ChatGPT, Codex, and Claude may have fed back into the companies’ own breakthroughs. The accusations are aimed at OpenAI and Anthropic, and the controversy flared again after OpenAI posted 377 more math results from an internal model that still hasn’t been released.

The sharpest fight came from Tristan Buckmaster, a New York University math professor working on the Navier-Stokes equations, one of the Clay Mathematics Institute’s Millennium Prize Problems. He says he and Levent Alpöge were close to publishing when Buckmaster also used OpenAI’s Codex to check his work. Then OpenAI announced in early September that its A.I. had solved the problem, and Buckmaster publicly said the company’s route looked far too similar to his own. OpenAI says 10,000 A.I. agents worked on the problem for about 88 hours. Buckmaster suspects his Codex sessions were accessed, and he says OpenAI researcher Sébastien Bubeck tried to push him to drop Alpöge as a co-author. Bubeck disputes that version but apologized for his wording.

A separate dispute centers on Andreas Thom of TU Dresden and a result OpenAI attributed to GPT-6 Astra in early August: a construction of a non-sofic group, a problem open for 27 years. Thom says the technical path used by the model matches work he and Gábor Kun have pursued for years, and he had discussed unpublished extensions with ChatGPT for months. He shared email exchanges with OpenAI staff, and when he asked whether those chats could have reached training data or been retrieved by the model, OpenAI researcher Mark Sellke replied that they had not. Thom says turning off training at the end of June would only stop future data, not erase what had already been shared. OpenAI later quietly revised its description of the result.

Anthropic is under fire for something similar. In late September, it said Claude, guided only roughly by in-house researchers, found a new enzyme system with CRISPR-like properties after about 950 A.I. agents searched genetic databases for 21 hours. Anthropic named the system Array-associated Reverse Transcriptase, or ART. Mario Rodríguez Mestre of the University of Copenhagen says he and his team had studied the same systems, which they call Jumbotrons, for about four years and identified them in 2022 without publishing yet. He says Claude had access to his dissertation manuscript, shared lab folders, and emails with collaborators. He has no proof, only a strong suspicion, but he plans to move to open-source models and is urging researchers to think twice before using proprietary ones.

OpenAI rejects the allegations and says neither researchers nor A.I. agents saw Buckmaster’s work before publication, and that chats were not accessed directly. But it also said it could not rule out that anonymized data from product use helped improve its models, before later saying Buckmaster’s Codex inputs could not have influenced the system at all. That uncertainty is exactly what worries scientists. If a chatbot can turn a whispered idea into a company demo, the incentive to keep quiet gets very real.