Mathematicians Urge OpenAI Boycott After 700 AI-Generated Papers Hit the Field

A group calling itself the Association for Human Mathematics is urging researchers to stop working with OpenAI, following the lab's release of more than 700 mathematics manuscripts generated by AI. The batch arrived all at once, and OpenAI pulled three of the papers back the next day after a sign error was found in them. The dispute has turned a technical question about proof verification into a fight over who gets to set the pace of mathematical publishing.

The scale is the sticking point. Seven hundred manuscripts is not a conference submission cycle; it is closer to the annual output of a mid-sized subfield, deposited in one move. Mathematicians who review for journals and referee preprint servers already describe overload, and a dump of that size shifts the burden onto human readers who had no say in the timing. The three retractions, meanwhile, gave critics a concrete example: an error that survived whatever checking preceded release.

Fields Medalist Terence Tao has framed the deeper concern in terms of what he calls harvesting. When AI systems mine open problems en masse, they consume questions that human researchers would otherwise spend years developing, and Tao warns that entire branches of mathematics can end up less fertile as a result. His argument is not that the answers are wrong, but that the pipeline treats unsolved problems as raw material rather than as a commons with a carrying capacity.

OpenAI has not said it will halt the practice, and no details have emerged on how the manuscripts were checked before publication or how the sign error was caught. That gap matters because the episode is being read as a test case for how AI labs release research artifacts into fields they are not part of. A retraction within 24 hours suggests some review happened, but not enough to prevent the error from shipping.

The boycott push also raises a question the mathematical community has dodged for two years: what obligations does a lab owe a discipline when it automates part of that discipline's work? Journals, preprint servers and funding bodies have no shared policy for AI-authored submissions at this volume, and the Association for Human Mathematics is effectively demanding one be written by refusing to cooperate until it is.

Watch whether the boycott gains signatories beyond the association's founding circle, and whether OpenAI publishes any protocol for how such batches are generated and screened. The retractions give both sides a factual anchor, and the next release, if there is one, will show whether the criticism changed anything.

For teams building on AI math systems, the practical lesson is verification cost: output volume is cheap, but adjudicating it is not, and any product that generates proofs at scale needs an internal error-catching step before release. Researchers should also expect journals and preprint servers to tighten AI-submission rules in the coming months, which will affect how generated results can be cited or claimed.