The Pulse
OpenAI’s 719 Math Manuscripts Stir Researchers’ Fears
OpenAI’s October 6 release of AI-generated mathematical work now contains 719 manuscripts grouped into 372 families, after three papers were withdrawn. Fields Medalist Hugo Duminil-Copin and other mathematicians say the release raises conce
AI.info Team · Written with AI ·
OpenAI withdrew three manuscripts from its mathematics release within a day of publication, after a sign error undermined one proof and work in two papers that depended on it. The company’s repository now lists 719 manuscripts grouped into 372 families of results, down from 722 manuscripts when the collection appeared on October 6. The correction has sharpened a dispute that reaches beyond whether individual claims hold up: mathematicians are also challenging how the results were produced, presented and credited.
Hugo Duminil-Copin, a mathematician who received the 2022 Fields Medal, said 16 problems in the release relate directly to his research. “They have annihilated my field of research,” he said during an October 8 videoconference organized by the Société Mathématique de France, according to Le Monde.
Three withdrawals put verification in focus
OpenAI’s repository history says a sign error invalidated an argument in one manuscript and the construction used by two related papers. The company withdrew all three on October 7. It also recorded revisions to 14 other manuscripts, including proof repairs, corrected statements and clearer hypotheses and dependencies.
The repository identifies the collection as work produced by an internal OpenAI model and says the results sit at different stages of verification. As of the current catalog, 300 of the 719 top-line results have Lean formalizations: computer-checkable versions of proofs written in a formal language. OpenAI says it will add formalizations as it obtains them, while warning that some results without them could have issues.
OpenAI’s October 6 announcement says each result required, on average, computing equivalent to about three hours of ChatGPT Pro thinking time. The company also says it evaluated the model on roughly 4,000 problems and released summaries of the model’s reasoning for 10 results. Those figures describe the company’s account of how the collection was generated; they do not establish that every manuscript’s argument is correct.
Advisers drew a line around endorsement
Before the release, the Advisory Group on Mathematics and Artificial Intelligence had published recommendations informed by more than 600 responses from the mathematical community. The independent group advised AI labs to stop testing advanced mathematical problems on proprietary models. It also called for careful sourcing, clear explanations of methods and computing costs, formal proofs where possible, and support for the human work needed to understand results that no researcher yet fully grasps.
After OpenAI published its collection, the group said its advisory role should not be mistaken for an endorsement or a judgment of the results’ impact. It described public release as the beginning—not the completion—of the work of understanding and incorporating the results into mathematics. “Mathematicians must be able to formulate their own questions, develop their own approaches, and explore directions that have not been selected as examples of an AI system’s capabilities,” the group wrote in its October 6 statement.
OpenAI said it consulted the group and was incorporating its recommendations. The tension is that the release came from a model OpenAI had not made available to outside researchers, while the group had specifically objected to frontier labs using proprietary systems to test advanced mathematical problems. The advisers left it to the mathematical community to assess how successfully their recommendations were followed.
Duminil-Copin sees a cost to open exchange
The volume of the release is part of the concern. Le Monde reports that one cluster connected to Duminil-Copin’s work included a 13-paper research program, and that he plans to confront more than 3,500 pages of generated material relevant to his field. He said reading the proofs did not have the same “flavor” as reading work by people whose faces he knew, but argued that mathematicians could still interpret, test and explain the findings together.
His sharper worry concerns what researchers may stop sharing. Duminil-Copin fears that companies could use online discussion to identify promising questions and deploy AI systems to pursue them, leaving human researchers with less opportunity for credit. He told Le Monde that some mathematicians had already decided to restrict their communications, and that “this entire culture of open exchange is being shattered.”
For now, the papers remain available for scrutiny, with OpenAI’s repository recording withdrawals, corrections and the status of formalized proofs. The immediate burden is concrete: mathematicians must assess a large body of work that the company says its own model produced, while deciding what kind of access, attribution and collaboration they expect around future releases.