New York University mathematician Tristan Buckmaster says he has stopped entering original ideas into OpenAI's Codex after its agents appeared to reach a stage of unpublished research he had developed with Anthropic mathematician Levent Alpöge. Buckmaster says the progress came within a day of OpenAI learning his identity. OpenAI denies that his prompts had any effect on its systems.
The dispute follows OpenAI's release of 722 AI-generated mathematics papers. Buckmaster also says model training was enabled by default on a personal account he used, raising questions about how unpublished research submitted to AI tools may be stored and used.
Buckmaster questions the timing of Codex's progress
Buckmaster and Alpöge spent nearly a year working on an approach to the Navier-Stokes problem, a longstanding challenge in fluid dynamics and one of the Clay Mathematics Institute's Millennium Prize Problems. A solution carries a $1 million prize. Buckmaster used a paid personal account to work on research drafts with Codex; the files were then stored on OpenAI's servers.
Buckmaster says the AI agents made no clear progress before OpenAI learned his name. Within a day of that, he says, the system had advanced to “exactly where we stopped.” He described the sequence as suspicious: “There was no progress before they knew my name; after they knew, they suddenly picked up where we stopped ... I find that suspicious.”
In September, Buckmaster said he was not accusing anyone. His view of the timeline changed after OpenAI later announced work on the Navier-Stokes problem, and he now describes the coincidence as suspicious.
OpenAI says Buckmaster's prompts had no effect
OpenAI has denied a connection between Buckmaster's data and its research. A company spokesperson said it could confirm that prompts Buckmaster submitted to Codex over the previous two months could not have affected the system in any way, including through model training.
Buckmaster argues that OpenAI's description of the relevant period changed. An earlier account referred to the preceding two months, while a later statement specified a start date of July 3. The differing time windows have become part of the disagreement over whether his data could have entered the system and how the reported research progress occurred.
Disagreement over collaboration and authorship
Buckmaster says that before OpenAI announced its Navier-Stokes research on September 8, the company offered to credit him as the sole author if he ended his collaboration with Alpöge. He declined and continued working with Alpöge.
The publicly stated facts are that Buckmaster used Codex to work on unpublished research, OpenAI denied that his prompts affected its systems, and the two sides differ over the timeline. For researchers and business users, the practical questions include whether model training is enabled by default, how long unpublished material remains on company servers, and how platforms distinguish user submissions from data used to improve models.