Twenty-five Fields Medalists — Terence Tao, Peter Scholze, and 2026 laureate Yu Deng among the initial signatories — put their names on a joint statement on September 11 titled “A Severe Misalignment of AI in Mathematics.” Le Monde’s headline version: the goals of the AI companies and the goals of the mathematical community are severely misaligned. And on the front page of Hacker News as I write this is Tao’s own Caltech lecture deck, “Math 2.0,” at 255 points and 312 comments. Same author, same season, two documents the internet has dutifully filed under “AI versus math.” Read them side by side and the fight is not about artificial intelligence at all. It is about grading.

The Statement Concedes Capability to Win the Argument

Nothing in the letter claims models can’t do mathematics. The signatories say the opposite — that large language models have improved dramatically, to the point of cracking problems that had sat unsolved for years. That concession is the entire story, and most coverage walked right past it, because “mathematicians say AI is bad at math” sells better than “mathematicians say AI is good at math, and that is precisely the danger.”

Their complaint is about selection. Once solving problems becomes the benchmark labs use to advertise capability, problems start getting chosen for machine-checkability rather than for interest. A theorem a verifier can confirm in an afternoon is worth more to a launch post than a theorem that reorganizes a field over twenty years. Reporting from 36Kr described the statement as the point where a priority dispute — who got credit, and for what — escalated from a release controversy into a collective condemnation of the method. The signatories include Pierre Deligne, a medalist from 1978. That is not a group confused about how mathematics works. It is a group that has watched the currency of its own field get reprinted by somebody else.

Benchmarks Are Procurement Documents

Here is the part that should bother anyone who trusts markets to sort this out. A benchmark is not a measurement. It is a procurement document. It tells funders what to pay for, graduate students what to learn, journals what to accept, and hiring committees what to respect. The moment a lab announces its model has resolved a famous problem, the announcement doesn’t merely describe a result — it re-prices every adjacent problem as either worth checking or probably not.

The Caltech slides sit at the top of Hacker News because Tao is one of the few people alive who can describe that repricing from the inside, and because he is honest enough to be interested rather than merely alarmed. At the 2026 International Congress of Mathematicians, per the Simons Foundation, he was already saying the field had to start asking philosophical questions about its own goals and values. That is not a man defending a monopoly. That is a man noticing that the thing being auctioned is the question, not the answer.

The Guild Is Defending the Right Turf, for Mixed Reasons

Let me be fair to the reflex on my side of the argument. Twenty-five medalists signing a joint letter is, functionally, a guild action, and guilds protect turf. Some of this letter is a jurisdictional claim wearing the clothes of a values claim, and the signatories would rather not say out loud that a machine is now a serious competitor for the prestige that keeps the seminar rooms full.

But the reflex picks the wrong turf. Mathematicians are not defending the ability to compute; the models have that, and no amount of letter-signing takes it back. They are defending the grading function — the human authority to decide which questions are worth asking. The genuinely uncomfortable implication for anyone who believes competition dissolves credentialism is this: competition is the mechanism doing the repricing. There is no market outcome where the labs stop buying referees, because referees are the cheapest input they have. A postdoc sitting on a fellowship committee put it to me in a Slack DM during application season: “We spent an hour Thursday arguing about whether a candidate’s thesis could be evaluated by a model, and about ninety seconds on whether anyone would still want to read it in twenty years.”

Every Other Discipline Is Eighteen Months Behind

Mathematics is simply the first field where the rubric is visible, because the rubric is proofs, and proofs can be checked by machine. Law, medicine, accounting, journalism, and the parts of software that aren’t typing are all somewhere behind it on the same curve, and they will hit the same wall the medalists hit in September: the work does not disappear first. The grading does.

That is what “misalignment” actually names. Not a technical defect that a cleverer training objective fixes, but a labor question about who owns the standard of proof and who gets paid when it changes hands. The twenty-five understood that, which is why they addressed the letter to each other and to their community, and not to a regulator. A regulator cannot grade a theorem. Only mathematicians can — and that is the asset everyone is now bidding on.

Sources