OpenAI posts 700 AI math preprints, withdraws three a day later

On 6 October, OpenAI released more than 700 AI-generated mathematics preprints to GitHub in a single batch. The manuscripts came from an unreleased model and claimed progress on 372 open problems in geometry, algebra and computer science, Nature reported. Within a day, three were withdrawn after a sign error invalidated an argument and broke the constructions in two dependent papers, and 14 others were revised, according to Retraction Watch. Three researchers in our Who's Who tracker had shared the Nature piece as the story spread.

The reaction from working mathematicians was sharp. "Why is OpenAI trying to solve hundreds of math problems to be released all at once?" asked Alvaro Lozano-Robledo, a number theorist at the University of Connecticut. Ursula Martin of the University of Oxford said the company was "tossing the community a messy first draft and expecting us to clean it up for them." Michael Harris, a number theorist at Columbia University, called the publication illegitimate and told Nature: "I can't bring myself to spend my time studying the outcome of this process...So I will not be reading these papers."

Not every reaction was hostile. Levent Alpöge, a mathematician at Anthropic, called the release "obviously the most significant moment in mathematical history" for showing how useful AI has become to working mathematicians.

Dan Roberts, an OpenAI research lead, said the company would "continue to update the repo with new formalizations and with any errata," and OpenAI told Nature it "welcomes scrutiny and feedback." Alex Townsend of Cornell said the company should have led with the subset formally verified in Lean, telling Retraction Watch: "They could have made a separate announcement for the other ones and ask for help." Andrew Sutherland of MIT called the withdrawal "the responsible thing to do" but said mathematicians remain "unhappy with how OpenAI has behaved up to this point."

None of the 722 manuscripts solved any of the five remaining Millennium Prize problems. OpenAI's 8 September release had tackled one of the original seven, Navier-Stokes.