A mathematician shared his ideas with AI. Now, he’s keeping his research away from OpenAI

A New York University mathematician has become more cautious about sharing original research with OpenAI after questioning whether ideas he entered into its Codex tool overlapped with the company's mathematical work.

Tristan Buckmaster has not shown that OpenAI used his private prompts. OpenAI says his Codex activity could not have influenced its results.

The case echoes a separate dispute involving biologist Mario Rodríguez Mestre, who said he shared unpublished research with Anthropic's Claude before the company announced an AI-assisted discovery in the same area.

Neither case has established that an AI company used a researcher's private work. However, together they raise questions about how scientists can use AI without risking ideas that have yet to be published.

The reason why the NYU mathematician stopped sharing his ideas with OpenAI

Buckmaster had used OpenAI's Codex, an AI coding agent, while exploring research ideas for solving a problem involving the Euler equation, which describes fluid motion. He had been working with Levent Alpöge, a mathematician at Anthropic.

On September 8, OpenAI announced a result involving the related Navier-Stokes equations. These equations describe how fluids, such as air and water, move. Buckmaster questioned whether his Codex conversations could have influenced OpenAI's work. However, the timing alone does not establish a connection.

While he continues to use the tools for other tasks, Buckmaster has since said he will keep his original ideas out of Codex to protect them. For researchers, an AI assistant may help test ideas, but sharing unfinished or unpublished work can be different, especially when the company running the service is also pursuing its own breakthroughs.

OpenAI denies private prompts influencing its research

OpenAI says it investigated and confirmed that Buckmaster's Codex prompts, submitted during the two months preceding its announcement, could not have influenced its research. It also says the proofs addressed different versions of the fluid equation, as Buckmaster and Alpöge worked on a case with an external force, while OpenAI's result concerned an unforced version.

In the case of the biology dispute, Anthropic said Claude identified a previously uncharacterized enzyme system. Mestre, however, says he and his colleagues had studied related enzymes for years. He had previously shared his dissertation and manuscript drafts with Claude.

OpenAI has also published mathematical results from an internal model, along with computer-checkable proofs for some of them. Such materials give researchers something to examine, but they do not settle every question about originality or credit.

There is no public evidence establishing that either AI company used the researchers' private work. However, this disagreement shows how hard it can be for outside researchers to access a company's internal development process. OpenAI has released supporting materials, but information related to its model and full development process is not open to independent review.

This dispute once again raises concerns about sharing unpublished data with AI

Neither case proves that an AI company took a scientist's idea. They do, however, show how hard it can be to access a proprietary AI system, especially when a company announces results that overlap with work still underway elsewhere. As AI moves beyond being a research tool, researchers should check the data policies and settings for the exact service they use before submitting their unpublished work.

OpenAI lets personal ChatGPT users opt out of model training. Anthropic says consumer chats may be used to improve Claude if users allow it or in certain other cases, while commercial offerings have different default terms.

Neither dispute establishes misuse of private prompts, but both show why scientists may keep their most valuable unfinished ideas out of AI chats until they are ready to make their work public.

Sources: The Washington Post, BeInCrypto

Affiliate links on Cult of Mac may earn us and our partners a commission.