OpenAI can't rule out training on a mathematician's notes

A 90-year-old math problem with a $1 million prize may have just been cracked by AI. There's a new question over who deserves credit.

OpenAI says a team of AI agents solved the Navier-Stokes problem, which asks whether the equations describing how fluids flow can break down.

But a mathematician racing toward the same answer says he wonders whether the company cribbed from the private notes he was feeding into its Codex coding tool, Wccftech reported.

Tristan Buckmaster of NYU and co-author Levent Alpöge posted a related proof the night before OpenAI's announcement. Leading mathematician Terence Tao wrote that their method looked likely to extend to Navier-Stokes.

The mathematicians were in communication with OpenAI staffers, and Buckmaster pressed them on whether the AI agents might have been trained on their work.

"I was told the model did not look up user data," Buckmaster wrote on Mastodon. "I asked again, about training, and I did not get an answer."

OpenAI said on X that no specific user data was accessed. But it added that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models."

That last sentence is the one to read twice: private work typed into an AI tool can quietly end up making the tool smarter.

DeAI angle: Lots of decentralized AI projects are working to provide confidential computing, where users can route requests to AI without worrying that sensitive data or solutions might be at risk.