25 Fields Medallists Tell AI Labs: Stop Using Our Problems as a Scoreboard

Twenty-five Fields Medallists published a public declaration on Thursday saying the goals of AI labs and the goals of the mathematical community are "severely misaligned," and warning that the industry's use of math problems as a capability benchmark is "detrimental to the science of mathematics." The letter, co-posted at mathandai.org, is the first coordinated objection of its kind from the credentialed core of the field, and its authors are explicit that math is one instance of a broader pattern.

The Fields Medal is mathematics' highest honor, awarded every four years to researchers under 40. The 25 initial signatories include Artur Avila, Manjul Bhargava, Pierre Deligne, Yu Deng, Simon Donaldson, Hugo Duminil-Copin, Martin Hairer, June Huh, Maxim Kontsevich, James Maynard, Peter Scholze, Terence Tao, Maryna Viazovska, Cédric Villani, Wendelin Werner, and Efim Zelmanov, among others. Tao released a supporting essay on his WordPress blog the same day explaining the declaration's reasoning and listing his own three harms in plain language.

The trigger was a claim from OpenAI three days earlier. The BBC and the Guardian both reported, on September 8, that OpenAI said it had used a large language model to crack a math problem that had stumped humans for decades. The OpenAI result is a company self-report, not an adjudicated result. The declaration treats the claim as a stand-in for a broader commercial pressure rather than a verdict on any specific paper.

What the signatories are objecting to is not the existence of AI tools. The mechanism, as Tao names it, is scoreboard capture. When an outside commercial benchmark starts measuring a field's progress, the field's internal progress metric, the one its practitioners use to decide what is worth working on, gets subordinated to the external scoreboard. The declaration's three stated harms follow from that. First, pushing math problems toward AI as a benchmark is "detrimental to the science of mathematics" because the problems most legible to a language model are not the problems that drive the discipline forward. Second, the mass production of true and false statements by AI systems "destroys the fertile ground for ideas," because a working mathematician's instinct for what is true depends on a slow, human-curated chain of conjecture, refutation, and revision. Third, the human transmission chain in math, the one that runs from graduate student to advisor to collaborator, is at risk when AI tools become the default interlocutor.

Tao compares the 2026 declaration to the 2025 Leiden declaration, an earlier, similarly structured statement on AI in research, and notes that this round was published with less consultation and under more urgency. The declaration's closing language extends the argument beyond math. The misalignment is, in the signatories' words, "part of broader alignment issues impacting other scientific and creative professions, as well as the whole of society." That sentence is what the signatories want other professions to read. It is a template.

The Hacker News thread on the declaration opened with a Baudelaire-on-photography parallel, the 19th-century poet's complaint that mechanical reproduction would destroy the fine arts. The Scientific American coverage uses a sharper word, "apocalypse," to describe the field's mood. Neither reaction is about whether AI can solve equations. Both are about professional values under commercial pressure.

The signatories have offered the letter. Whether the OpenAI result is a real mathematical advance or a benchmark-solve will test the scoreboard-capture argument. The Economist has run a companion piece on the declaration, referenced in Tao's essay. Other fields, from physics to biology to computer science itself, will have to decide whether the template fits. Math is the test case the signatories chose, and the one they expect other professions to read.