OpenAI Releases 722 AI-Generated Math Manuscripts After Criticism

OpenAI published 722 AI-generated mathematical manuscripts on GitHub at 6 p.m. ET Oct. 6 (22:00 UTC), putting a large body of machine-produced research into public view while drawing criticism over how the work was disclosed.

The manuscripts are organized into 372 related papers. Many contain Lean formalizations, which allow mathematical work to be checked by computer, although not every manuscript has been formalized. OpenAI’s README acknowledged that some nonformalized results might contain problems and said they would be fixed as soon as possible.

The release follows OpenAI’s claim that a new model solved more than 100 long-standing problems in addition to the Navier-Stokes problem, which concerns equations describing fluid motion. OpenAI previously said it used more than 10,000 AI agents on Navier-Stokes and estimated the computing cost at several million dollars.

OpenAI said the average result represented about three hours of ChatGPT Pro computation. OpenAI included average computing requirements and 10 reasoning summaries, but did not publish the prompt used for each problem. The company also said nearly every result came from a single prompt given to a single AI agent.

That approach differed from recommendations issued Sept. 29 by the Advisory Group on Mathematics and AI, or AGMAI. The nine-member group is established under the Institute for Advanced Study, based in Princeton, and its members do not receive payment for the work. Its recommendations drew on responses from more than 600 mathematicians.

AGMAI recommended placing results in an academic repository not controlled by an AI laboratory. It also called for the model name, prompts, reasoning summary, time spent and computing cost to be disclosed for every result. The group said mathematical work should not be used as a marketing tool.

OpenAI’s manuscripts were placed in the company’s GitHub account. The release also said OpenAI was looking for a community-managed repository that would meet the group’s recommendations.

The disclosure followed an August meeting at which mathematicians discussed how AI-generated mathematical results should be published and reviewed. Bryna Kra, a Northwestern University mathematician who attended, said participants urged OpenAI not to rely only on blogs or posts and instead to write papers that mathematicians could assess and use.

Andrew Sutherland, an MIT mathematician, said claims about the results should be treated as unverified until the model is made public and the work can be reproduced. AGMAI said its advisory role did not endorse the manuscripts or OpenAI’s process and that publication was only the beginning of review.

Daniel Litt of the University of Toronto viewed public disclosure more favorably, saying companies had no reason to keep mathematical answers secret. Kra helped draft the Leiden Declaration, which has been signed by more than 4,000 mathematicians and calls for AI companies to meet standards established by the mathematics community.

The release adds to OpenAI’s earlier disclosure of a Navier-Stokes proof, as mathematicians begin assessing whether the new results can be independently reproduced and incorporated into formal research.