25 Fields Medal Winners Sign Open Letter Warning AI Labs Are Threatening Mathematicians' Intellectual Labor
Twenty-five Fields Medal winners have signed an open letter stating that AI labs are threatening mathematicians' intellectual labor by rushing to announce solutions to famous mathematical problems. Each signatory has received the highest award in the field.
The letter states: Solutions are often announced hastily, leaving no time to write complete papers, separate new methods, and cite previous work; this raises issues of attribution and plagiarism in all creative professions. Without mathematicians willing to write new ideas into mathematical classics and complete the human teaching chain, ideas conceived by AI cannot truly come to life. The signatories believe that if models can solve outstanding problems, it can benefit humanity, provided that the solutions can be understood and disseminated by the mathematical community. Public reports mention signatories including Terence Tao, Artur Avila, and Manjul Bhargava.
The trigger was a priority conflict that occurred in the same week. NYU professor Tristan Buckmaster accused OpenAI of pressuring him not to thank a collaborator at Anthropic and questioned whether the company was leveraging their work with Codex to produce its own breakthrough proof through marathon reasoning. OpenAI subsequently published a complete proof regarding the Navier-Stokes existence and smoothness Millennium Prize problem, claiming it was completed by an unpublished next-generation model in about a week, consuming approximately 300 billion output tokens. Buckmaster's team had previously announced three proofs advancing towards this problem. OpenAI also withdrew sponsorship after facing criticism from campus researchers during a mathematics event at Caltech. In June, the Leiden Declaration had already called for mathematicians, institutions, and governments to regulate the use of AI in research, receiving endorsements from the International Mathematical Union and others.
Labs treat unverified proofs as advertisements for their capabilities; mathematicians view unfinished manuscripts and citations as conditions for the survival of the discipline. Each Millennium Prize problem has a reward of $1 million, but the open letter contests who has the right to announce and who must write the methods into verifiable texts.
In market mechanisms, this is a struggle for reputation and narrative training, not prize settlement. The buyers are model companies announcing "solved" on social platforms; the sellers are the mathematical community that still controls peer review, citations, and classroom teaching. The beneficiaries are labs that can turn famous conjectures into evaluation and financing stories; the pressured are scholars who must verify after hot announcements but do not receive computational budget. Funds flow from computational bills to model release events, while verification costs are left to universities.
Supplementary structure: OpenAI's proof is still described in the letter as insufficiently verified. There is also a joint statement criticizing the use of large models to generate unverifiable mathematical garbage for students at Caltech's "Mathematics Marathon."
Source: Public Information
ABAB AI Insight
The collective letter from Fields Medal winners is not against machines solving problems, but against turning mathematics into a race for publication. For Millennium problems like Navier-Stokes, the traditional path involves years of writing and years of peer review; the lab's path is one week, 300 billion tokens, and a post first. Buckmaster and his collaborator at Anthropic being pushed to the edge of the acknowledgment list illustrates how computational advantages can rewrite academic authorship. The Leiden Declaration set out values in June, and this letter in September specifies the conflicts with concrete personnel.
The capital path is that evaluation equals market value. A proof, even if not written into the classics, can add value to the next generation of models. 300 billion tokens, at public reasoning prices, can be equated to tens of millions of dollars in computational costs; this money buys a timestamp of "we were first," not a check from the Clay Institute. What universities lose is the window to teach methods to the next generation: without writing, there are no theorems to teach. The withdrawal of sponsorship from Caltech indicates that labs have discovered the mathematical community is beginning to replace acknowledgments with resistance.
This is analogous to pharmaceutical companies rushing to release incomplete Phase III data and the patent rush in early gene sequencing: the social function of scientific discovery is reproducible texts, while the market function is monopolistic narratives. The industry phase is that cutting-edge models treat formal proofs as a capability arms race, while mathematicians view formal proofs as literature that must be understood by humans. Both use the same terms, but the currencies of settlement differ.
The structural judgment shifts pricing power from the peer review community to computational timestamps. The mechanism is: the correctness of mathematics is binary, but "who says it first" is continuous on social media. Labs optimize announcement speed, while the community optimizes citability. Without someone willing to write into the classics, machine solutions remain in press releases, but training data continues to treat them as solved.
ABAB News · Cognitive Laws
Announced proofs do not equal proofs written into the discipline.
Computational power can seize timestamps, but cannot take away the teaching aspect in classrooms.
Once creative industries only compete on speed, ownership rights will bankrupt before correctness.