On September 11, Terence Tao published a short declaration on his blog and pointed readers to mathandai.org. It is titled “A Severe Misalignment of AI in Mathematics.” It carries 25 signatures, every one of them a Fields Medalist. Within a day it sat at the top of Hacker News with 928 points and 890 comments — roughly the ratio of people who can read a proof to people who have an opinion about one.
Here is what the signatories actually say, which is not what the reaction assumed. They do not claim the machines can’t do mathematics. They claim the opposite: that over the last few months, LLM mathematical capability has improved dramatically, to the point where these systems can solve major outstanding problems across many fields. Their objection is narrower and stranger. Benchmark-driven AI, they argue, races through famous open problems without crediting or involving the mathematicians who posed and nurtured them; mass-produced solutions arrive without writeups; and the intermediate reasoning that once turned a result into transmitted knowledge gets skipped.
That is not a complaint about capability. It is a complaint about the audit trail. And it is the part of the story that both the guild-defenders and the lab-defenders have talked straight past.
The Ask Is Attribution, Which Is a Property Right
For roughly twenty-five years, the respectable free-market position on information was that attribution norms were a nuisance — a tax on remixing, a lawyer’s full-employment program, something the music industry whined about while its business model died. Silicon Valley built an entire jurisprudence on the premise that transforming an input produces a new thing that owes nothing to the old one.
Now 25 of the most credentialed mathematicians alive are asking for precisely that tax. The reflex — this is a guild protecting its turf — gets it half right and then stops thinking. What is being defended is not prestige. It is the chain of custody. A result that cannot be traced through its lemmas is not a theorem; it is a press release with notation. The claim circulating this month, that OpenAI’s internal model produced a Navier-Stokes proof, with the company’s ambitions now pointed at the Riemann Hypothesis and P versus NP, arrives as an assertion rather than a verified artifact. Structurally, that is the same event as a company announcing earnings with no auditor’s signature and no ledger. Everyone understands why that isn’t allowed in finance.
One engineer at a frontier lab, answering a direct message well after midnight on the day of the declaration, put it without much spin: “We don’t have a proof. We have a model that says it’s a proof and a room full of people who need it to be one.”
The Benchmark Was Their Idea
Here is the part of the declaration its own signatories half-admit, complaining about “the push by AI companies to solve mathematical problems as a benchmark.” The labs did not invent benchmark culture. They inherited it. Mathematics spent decades making itself legible to outsiders as a leaderboard — named open problems, prize committees, a pipeline that sorts teenagers by timed olympiad performance, a public conversation in which “solved” is binary and “why” is decoration. That was a choice, made long before anyone had a transformer, and it had real benefits: money, attention, a legible case for funding.
It also made mathematics the one field where a machine’s output can be checked in an afternoon. Which brings us to the actual misalignment, and it is not between AI and mathematics.
Mathematics Is the Only Field That Could Have Noticed
Radiology, litigation, tax accounting, contract review, the trade press, whatever this column is — every one of them has spent three years absorbing unreferenced machine output and shipping it. Not one produced a declaration signed by 25 of its most decorated practitioners, because not one of them can reliably tell a good output from a fluent one. Mathematics filed its protest within days, not because mathematicians are braver, but because a proof either verifies or it doesn’t. That single property gave them the detection mechanism and the standing to complain.
The uncomfortable implication runs in both directions. If your instinct is to let the labs cook, you owe an answer to why the audit function is not part of the product — and “the market will sort it out” is not an answer when the buyer cannot evaluate the good he is buying. If you are a mathematician, you have to concede that you spent decades selling your field as a scoreboard and are now startled that the score is being kept by someone else, in a ledger you don’t control.
The misalignment is not that AI is badly suited to mathematics. It is that mathematics is the only discipline left built with a tripwire in it. Everyone else absorbed the same thing quietly and called it productivity.