Mathematicians warn AI proof race could undermine attribution
Twenty-five Fields Medal recipients have signed an open letter warning that competition among AI laboratories to solve celebrated mathematics problems could threaten attribution, verification and open research. The intervention follows allegations involving OpenAI, an unverified proof associated with the company, and its withdrawal of sponsorship for a mathematics event at Caltech.
The signatories argue that AI-generated solutions may benefit humanity only if the mathematical community can understand, communicate and develop them. They warn that announcing results too quickly can leave insufficient time for a proper write-up, identification of new methods and ideas, and citation of relevant earlier work.
Concerns over credit and verification
NYU professor Tristan Buckmaster accused OpenAI of pressuring him not to credit a collaborator employed by Anthropic for solving an important problem. He also questioned whether work shared through Codex may have been used to produce OpenAI’s claimed proof during a marathon inference run. The proof has not been verified.
The letter frames the issue as more than a contest to produce a correct answer. Mathematicians must assess proofs, explain their methods, connect them with existing knowledge and carry those ideas into teaching and further research. Without that work, the authors say, AI-conceived ideas cannot become fully integrated into the mathematical canon.
Pressure on open research
Researchers are also concerned that their use of Codex could feed into newer OpenAI models. In that environment, the authors fear a laboratory that identifies a promising route to a discovery could spend tens of millions of dollars on large language model inference and reach a proof ahead of the original researchers. They argue that such a dynamic could encourage secrecy rather than collaboration.
The warning follows the Leiden Declaration, issued in June by a working group of mathematicians. Its recommendations address how mathematicians, institutions and policymakers should respond as large language models alter mathematical practice. A related dispute over AI evidence and records appears in OpenAI litigation over logs and evidence as questions about transparency and accountability extend beyond research credit.
Implications beyond mathematics
The signatories compare the challenge with changes already affecting software engineering and other professions. Their central concern is that the value of mathematics includes the surrounding intellectual structure: developing students, posing new questions and transmitting knowledge between people, not merely producing proofs or assigning credit.
For organisations using AI in expert workflows, the practical implication is to preserve review, provenance and attribution processes so that useful output can be checked, explained and responsibly incorporated into human work.