Twenty-five Fields Medal winners warn of misalignment between AI and mathematics

Twenty-five Fields Medal winners, including Terence Tao, have signed a joint declaration warning that AI companies and the mathematics community are fundamentally misaligned. The statement, published on September 11, 2026, argues that the commercial incentives driving AI development are clashing with the rigorous, slow-by-design processes that have kept mathematics reliable for centuries.

What the declaration actually says

The core argument is deceptively simple. AI companies are optimizing for speed, claiming breakthroughs on mathematical problems at a pace that outstrips the community’s ability to verify them. Mathematics, unlike most fields, has a binary quality standard: a proof is either correct or it isn’t.

The signatories, who include Artur Avila and Manjul Bhargava alongside Tao, frame this as an alignment problem. Not the existential AI alignment that dominates Silicon Valley discourse, but a more immediate, practical one. AI firms want impressive results on competitive timelines. Mathematicians want results that are actually true.

The declaration points to recent incidents where AI systems have made claims related to long-standing mathematical problems, including problems in the Millennium Prize category. These are the seven hardest unsolved problems in mathematics, each carrying a million-dollar bounty from the Clay Mathematics Institute. When an AI company announces progress on one of these problems, it generates enormous publicity. When the community later finds the work doesn’t hold up, the retraction gets considerably less attention.

The mathematicians argue this creates a dangerous cycle. Flashy claims attract funding and talent to AI math projects, which produces more flashy claims, while the painstaking work of verification falls to an academic community that doesn’t operate on quarterly earnings timelines.

The Leiden Declaration and a growing pattern

This isn’t the first time the academic world has tried to draw boundaries around AI’s encroachment into expert domains. The new declaration follows the Leiden Declaration from June 2026, which addressed broader risks that AI poses to professions built on deep human expertise.

The Leiden Declaration identified several specific threats, including unreliable proofs and citation issues. If that sounds abstract, consider the practical version: an AI system generates a proof that looks correct on the surface but contains subtle errors that only a specialist would catch. If that proof gets cited by other researchers, or worse, used as a building block for subsequent work, the error propagates.

The Fields Medalists’ statement builds on this foundation but sharpens the focus. Where the Leiden Declaration painted with a broad brush across multiple professions, this new letter zeroes in on how AI companies’ market-driven timelines specifically undermine mathematical integrity. The signatories reportedly spent the week before publication in discussions refining their position, suggesting this wasn’t a hasty response to a single incident but a considered assessment of a systemic trend.

Tao himself has been one of the more nuanced voices on AI in mathematics. He has previously engaged with AI tools and acknowledged their potential utility in mathematical research. That makes his willingness to lead this declaration more notable.

Why the incentive mismatch matters

The declaration frames the gap not as a temporary growing pain but as a fundamental misalignment in values. Deep understanding, proper attribution, and thorough verification aren’t bureaucratic overhead in mathematics. They are mathematics. Strip them away in the name of speed, and what remains may look impressive in a press release but fails the discipline’s actual standards.

Andrew Wiles spent seven years working on Fermat’s Last Theorem in near-total secrecy. The initial proof contained an error that took another year to fix. The mathematical community considered this entirely normal.

For now, the declaration is exactly that: a declaration. It carries moral authority but no enforcement mechanism. AI companies are under no obligation to slow their claims or submit to community review processes. But 25 Fields Medalists speaking in unison creates a reference point that policymakers, journal editors, and funding agencies can point to when making their own decisions about how to handle AI-generated mathematical claims.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.