Tristan Buckmaster, a mathematician at NYU, claims that an OpenAI model reproduced his unpublished research on the Navier-Stokes fluid-flow puzzle within one day of identifying his name. Buckmaster and Levent Alpöge from Anthropic spent nearly a year working on the problem, which carries a $1 million prize. The researcher processed his drafts using the OpenAI coding assistant Codex via a paid personal account.
Buckmaster noted that the model showed no progress on the specific problem until he provided his identity, after which it reached the same conclusions as his team. He also stated that OpenAI proposed he take sole authorship of the proof if he excluded his collaborator, an offer he declined. OpenAI officially denied that Buckmaster's prompts influenced its systems, stating that such an impact is impossible.
The company initially claimed that prompts from the last two months could not have affected its training, but later adjusted this timeframe to July 3. Buckmaster now advises researchers against inputting original ideas into AI tools. He highlighted that grant reviewers often use these systems to summarize proposals, potentially exposing sensitive research to AI firms.
OpenAI recently released 722 math papers generated by its AI, which drew from 4,000 unsolved problems. According to Buckmaster, this development negatively impacted the research programs of early-career mathematicians. Meanwhile, OpenAI stated that it has achieved substantial progress on another problem associated with the $1 million prize.
