25 Fields Medalists Sign 'A Severe Misalignment of AI in Mathematics': The Discipline's Highest Honors Formally Break With the Labs

On September 11, 2026, something unprecedented appeared on Terence Tao’s blog. Under the title A Severe Misalignment of AI in Mathematics, a single declaration carried the signatures of 25 Fields Medal winners — Avila, Bhargava, Birkar, Deligne, Deng, Donaldson, Duminil-Copin, Figalli, Hairer, Huh, Kontsevich, Lindenstrauss, Lions, Maynard, McMullen, Mori, Ngô, Okounkov, Scholze, Smirnov, Tao, Viazovska, Villani, Werner, and Zelmanov. For those keeping score, that is a meaningful fraction of every living Fields Medalist on Earth, spanning medal years from 1978 (Pierre Deligne) to 2026 (Yu Deng).

The timing was not accidental. It came three days after OpenAI’s contested Navier–Stokes announcement triggered a priority dispute with NYU’s Tristan Buckmaster, two days after an open letter from 771 mathematicians forced OpenAI to withdraw sponsorship of Caltech’s AI-driven “Mathathon,” and roughly two months into a season in which AI systems have been knocking over famous unsolved problems — the Erdős corpus, the Riemann sphere’s peripheral conjectures, frontier benchmarks — at a pace the field has never seen. The declaration is mathematics formally drawing a line: not against AI, but against what the labs have made of it.

What the Declaration Actually Says

The text is short — under 800 words — and measured in tone, which makes its central sentence hit harder: “The push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics, and to the mathematical community. The goals of the AI companies and the goals of the mathematical community are severely misaligned.”

The argument proceeds in four moves.

First, famous problems are infrastructure, not trophies. The declaration describes landmark conjectures as “landmarks and lighthouses against which one can measure an improved understanding of this landscape.” Solving one has historically been a certain sign of new insights and methods, which the community then absorbs through “a long and arduous process of talks, discussions, simplifications” — ending, ideally, in a textbook presentation a graduate student can study. Some of those ideas travel further still, becoming tools “understood and used by the whole population” decades or centuries later. The value was never the answer. It was everything the attempts generated along the way.

Second, AI labs are now optimizing against that value. LLMs can “solve major outstanding problems in many fields of mathematics” — the signatories concede this without hesitation. But “solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight. Forgetting this in the world of AI may turn the tool against the primary goal.” The mass production of true/false statements “at faster and faster pace” could, in their words, “destroy fertile ground instead of breathing life into new ideas.”

Third, the announcement culture is the immediate injury. Solutions arrive “in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others.” That raises “severe attribution and plagiarism questions” — and it externalizes costs onto working mathematicians, who must verify, contextualize, or discredit claims they had no part in, labor that is uncompensated and unacknowledged. And without “willing mathematicians” to develop and integrate AI-conceived ideas into the canon, “the crucial human transmission chain between mathematicians would be lost.”

Fourth, mathematics is the canary, not the exception. The declaration explicitly frames this as “a general threat to intellectual work” — the same misalignment that other scientific and creative professions face, and “issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place.”

Notably, the closing paragraphs are not anti-AI. “AI offers the potential of enhancing and accelerating genuine mathematical study and understanding,” the text reads, and the profession “will need to adapt.” The question the signatories pose is who decides: “whether these changes ultimately benefit the field or have a destructive effect will in large part be determined by the decisions of the humans in control of this new technology.”

How We Got Here in Ten Weeks

The declaration did not emerge from a vacuum. It is the crest of a wave that has been building since early July, when an Erdős-problem gold rush first made AI mathematics a mainstream story. Since then the escalation has been relentless:

• July–August: AI-assisted solutions to long-standing problems start arriving weekly, often announced on social media with minimal writeups — the pattern the 771-mathematician letter would later call “slop mathematics.”
• September 8: OpenAI’s Navier–Stokes regularity claim lands amid a priority dispute with Buckmaster, who had been working the same terrain. Tao’s response essay argued the incident showed the danger of autonomous proof-hunting with no human insight extracted.
• September 10: The Mathathon open letter (771 signatories, now past a thousand) calls the Caltech event’s model “destructive” and labels the labs’ behavior “research misconduct.” OpenAI withdraws its sponsorship the same day.
• September 11: The Fields Medalists’ declaration lands — explicitly noted by Tao as the product of “discussions between ourselves over the last week,” rushed out without the consultative process of the Leiden Declaration because “the urgency of the situation was such that we needed to release a statement sooner rather than later.”

Read in sequence, the pattern is clear: the community’s institutional response has escalated from individual essays (Tao, early September), to mass grassroots letters (771 researchers), to now the field’s most decorated individuals speaking as a bloc. The Economist’s coverage framed it bluntly: “Top mathematicians are outraged by OpenAI’s methods.”

Why This Time Is Different

Skeptics will note that mathematicians have complained about computers for decades, and that every previous alarm — computer-assisted proofs like the Four Color Theorem in 1976, or experimental mathematics in the 2000s — ended with the field absorbing the new tools. Three things make this declaration different.

The signatories are not skeptics of AI. Tao has been among the technology’s most thoughtful optimists, formalizing proofs in Lean and writing extensively about AI as a collaborator. Scholze built his career on exactly the kind of conceptual restructuring that AI proof-mining bypasses. These are people who have spent the last two years using these systems, and their objection is not capability but incentive design.

The complaint is economic, not romantic. The declaration’s core claim — that benchmark-driven problem-solving extracts a non-renewable common resource (good problems, and the careers built on them) without replenishing the ecosystem — is a precise argument about externalities, not nostalgia. It mirrors the argument Tao made in his “drinking water beside an ocean” essay: the scarce resource in mathematics was never answers.

It names a governance vacuum. By connecting mathematics to “other scientific and creative professions” and “the whole of society,” the declaration is effectively asking for settlement of a question the labs have deferred: who bears responsibility for the commons when the tools of extraction are privately controlled? That is a policy question, and it is now formally on the table — the same week the U.S. Senate was debating a duty-of-care regime for frontier AI companies.

What Happens Next

The declaration invites further signatures at mathandai.org, and if the Mathathon letter’s trajectory (771 to 1,000+ in days) is any guide, the list will grow. The labs now face an awkward sequence: OpenAI has already retreated from the Mathathon, but the declaration’s ask is far bigger than one hackathon — it is a demand that the companies stop treating famous problems as free benchmark fuel and start investing in the human infrastructure of verification, exposition, and attribution.

There are early signs of what compliance could look like: formal verification pipelines that make AI proofs checkable, publication norms that credit concurrent human work, funding for the mathematicians who integrate AI output into the canon. Whether any lab adopts them voluntarily is another matter. As the declaration’s final sentence makes clear, the signatories believe the window is short: “These issues must be addressed urgently.”

For a field that usually measures time in decades, mathematics has moved with remarkable speed this month. The question the labs must now answer is whether the fastest-solving machines in history can also learn to slow down.