25 Fields Medallists Rushed Out a Letter About Rushing

TL;DR

On September 11, twenty-five Fields Medallists published A Severe Misalignment of AI in Mathematics, a short declaration arguing that the push by AI companies to solve famous mathematical problems as a benchmark is actively damaging the field. The signatories span award years 1978 to 2026. Terence Tao, who is one of them, says the text grew out of discussions "over the last week." It names no company. As of this writing on September 12 it has 1,694 endorsers, and the counter is still moving.

The argument is Goodhart's law, applied to a whole science

The declaration's core move is to separate the metric from the thing the metric was measuring. Famous open problems, it says, "have often served as landmarks and lighthouses against which one can measure an improved understanding of this landscape." Solving one used to be strong evidence that somebody had found a genuinely new idea, because there was no other way to get there.

That inference is what breaks. Once a system can produce the solution without producing the understanding, the landmark stops certifying anything. In the declaration's words: "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."

Think of the famous problems as summit cairns on a mountain range. They were never valuable because somebody stood on top of them. They were valuable because the route somebody took to get up there became a path other people could walk. A helicopter that drops a flag on every summit has produced a great many flags and exactly zero trail.

The second half of the complaint is procedural, and sharper. "Often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions."

Four days from claim to declaration

This did not come out of nowhere. The week it responds to was unusually loud.

On September 8, OpenAI published a claimed resolution of one direction of the Navier-Stokes Millennium Prize Problem, with Lean certificates, hours after the two mathematicians working the same route posted a statement about how that had gone for them. On September 9, TU Dresden group theorist Andreas Thom published his email exchange with OpenAI researchers, asking whether months of his unpublished conversations with ChatGPT had reached the training pipeline. He got an answer to a narrower question than the one he asked.

Two days later, the declaration. The Economist ran a piece on it the same day.

Who actually signed

Twenty-five medallists, from Pierre Deligne (1978) to Yu Deng (2026). The middle of the distribution is where the field's current centre of gravity sits.

The list includes Tao, Peter Scholze, Maryna Viazovska, James Maynard, Maxim Kontsevich, Cedric Villani and Pierre-Louis Lions. Endorsement is gated behind ORCID sign-in or a confirmed academic email, which is a polite way of saying the bots are not invited.

The speed is the tell

Mathematics already had a declaration about AI. The Leiden Declaration on Artificial Intelligence and Mathematics was published on 2 June 2026 with a Zenodo DOI, it is endorsed by the International Mathematical Union, and it grew out of a September 2025 workshop at the Lorentz Center in Leiden. Nine months of process, institutional backing, the full apparatus.

This one took a week. Tao says so directly: "It is unfortunate that we did not have the time to have a more consultative process, as with Leiden; but we decided that the urgency of the situation was such that we needed to release a statement sooner rather than later." A declaration whose central grievance is that results get announced in a rush, announced in a rush, with the apology in a parenthetical. That is about as close as a Fields Medallist gets to a shrug.

Leiden took 102 days to reach 3,915 signatories. The new declaration reached 1,694 endorsers in roughly one. Different mechanisms and different words on the button, so do not read this as a clean rate comparison, but the direction is not subtle.

The attribution complaint already has a receipt

The declaration's line about "citing relevant previous work of others" is not hypothetical. In August, OpenAI's ten mathematics results included the first construction of a non-sofic group, settling a question Mikhail Gromov posed roughly 27 years earlier. The construction leans on a 2019 expander-graph theorem by Gabor Kun and Andreas Thom. OpenAI's writeup was quietly updated to credit them after mathematicians pointed out the omission.

That is also why Thom's question lands harder than a generic privacy complaint. He had spent months discussing adjacent unpublished problems with ChatGPT. The reply he received addressed whether the system read his conversations while solving, which is a different question from whether they entered training. A follow-up paper by Francesco Fournier-Facio has since produced torsion-free examples from the same criterion, which is the normal machinery of mathematics doing what the declaration says it needs time to do.

Why this should matter to you even if you never open Lean

Three things generalise straight out of mathematics and into anything you ship.

Your eval is not your deliverable. "The tests pass" was a good proxy for "the code is correct" precisely while writing code that passes tests was hard. Point a capable agent at the proxy and the correlation is the first thing to go. This is Goodhart's law with a Fields Medal attached, and it is the most portable idea in the document.

Emit citations or inherit the plagiarism problem. The most decorated people in a field just put attribution and plagiarism in a joint statement, with a concrete case of a writeup being corrected after the fact. If your pipeline synthesises from prior art, make it name the prior art in the artifact, not in a footnote you add when somebody complains.

Unpublished work in a hosted model is an unfalsifiable bet. The Thom episode is instructive less for what OpenAI did than for what nobody outside can establish. Even a truthful denial is unverifiable from where you sit. If your novel work is the whole asset, the threat model is not "will it leak," it is "could you ever prove it did not."

What the declaration does not do

It is a position statement, and it carries no enforcement, no institutional endorsement, and no request more specific than "these issues must be addressed urgently." It names no company, so the Economist headline about OpenAI is an inference from context rather than a quotation.

The 25 are also a self-selected group assembled in a week, not a poll of living medallists, and plenty of prominent mathematicians are not on the list. The endorser count is a live, unaudited number on a site that went up yesterday. And the declaration does not dispute that the AI results are correct. Its whole premise is that they are, which is what makes the complaint interesting rather than defensive.

Key Takeaways

• Twenty-five Fields Medallists, spanning award years 1978 through 2026, signed A Severe Misalignment of AI in Mathematics on September 11, 2026.
• The argument is Goodhart's law at field scale: solving famous problems was only ever a proxy for conceptual understanding, and optimising the proxy can destroy what it measured.
• The secondary complaint is procedural, covering rushed announcements, missing writeups, and attribution, with OpenAI's quietly corrected non-sofic groups credit as a live example.
• The Leiden Declaration took nine months from workshop to publication and has 3,915 signatories in 102 days. This one took a week and drew 1,694 endorsers in about a day.
• It names no company, has no institutional backing, and asks for nothing specific. Its force is entirely in the signatures.
• The transferable lesson for builders: when the benchmark becomes the deliverable, the benchmark stops telling you anything, and attribution is the first thing to fall out of a fast pipeline.

Sources: A Severe Misalignment of AI in Mathematics (mathandai.org), the endorser list and count, Terence Tao, "A Severe Misalignment of AI in Mathematics", Leiden Declaration on Artificial Intelligence and Mathematics, Leiden signatory list, Leiden Declaration DOI 10.5281/zenodo.20302944, Wikipedia on the Leiden Declaration's origins, The Economist, "Top mathematicians are outraged by OpenAI's methods", Andreas Thom on Mathstodon, OpenAI, "Ten advances in mathematics and theoretical computer science", OpenAI on Navier-Stokes, Tristan Buckmaster and Levent Alpoge, statement (PDF), Francesco Fournier-Facio, "A torsion-free non-sofic group" (arXiv 2608.02025), Terence Tao, "On the existence of non-sofic groups"