NYU Mathematician Says OpenAI Model Matched His Unpublished Results Within a Day

A mathematician at NYU says an OpenAI model matched his unpublished findings within a day after it learned his identity, and he now withholds original research ideas from the firm’s AI products. OpenAI maintains that his input could not have influenced its system.

Buckmaster issued that caution just days after OpenAI put out 722 math papers written by AI. One of his own accounts, he said, was still set to allow model training by default.

What Buckmaster Says Happened to His Unpublished Work

Alongside Levent Alpöge, a mathematician at OpenAI competitor Anthropic, Buckmaster devoted the better part of a year to a piece of the Navier-Stokes problem, a fluid-flow challenge carrying a $1 million prize.

He ran the work through OpenAI’s coding assistant Codex on a paid personal account, and the drafts sat on the company’s servers.

In an interview with CNBC, Buckmaster said OpenAI had no progress to show before he disclosed his name. Within a day of that disclosure, he added, the company’s agents reached “exactly where we were.”

BeInCrypto reported in September that Buckmaster said he was not accusing anyone. Now he describes the timing as suspicious.

“Before they knew my name, they had no progress. After they knew my name, they suddenly started from where we left off…I find it suspicious,” Buckmaster said.

Buckmaster says that ahead of OpenAI’s September 8 announcement of its Navier-Stokes proof, the company offered him sole authorship on condition that he drop Alpöge. He declined.

OpenAI dismisses the idea that his data had any connection to its work.

“We can say categorically that it is impossible for Dr Buckmaster’s Codex prompts over the last two months to have influenced the system in any way, including training,” ABC reported, attributing the comment to an OpenAI spokesperson.

Buckmaster says that time frame kept changing, and a later company statement placed it at July 3.

Why Buckmaster Now Tells Researchers to Keep Ideas Out of AI

Buckmaster still uses Codex, but he no longer feeds in ideas he wants to remain original.

Grant reviewers, he says, frequently paste research proposals into AI tools to produce summaries, which passes unpublished plans on to AI companies.

OpenAI’s October release drew on 4,000 unsolved problems. Buckmaster said it erased the research programs of early-career mathematicians.

Sébastien Bubeck, OpenAI’s lead researcher, told The New York Times his offers were intended to help Buckmaster.

A spokesperson said OpenAI has made “substantial progress” on a different one of the $1 million problems.