On September 11, Terence Tao — UCLA, Fields Medal, the closest thing mathematics has to a household name — posted a declaration titled “A Severe Misalignment of AI in Mathematics,” hosted at mathandai.org and initially signed by 25 Fields Medalists. It grew out of a week of discussions among the signatories, as the journalist Kevin Hartnett noted. Two days earlier, per the Chosun, Tao had complained that the labs handle only the “initial stages of discovery and verification” and walk away from the rest. One day before that, OpenAI announced a solution to the Navier–Stokes problem — a puzzle that has stood for ninety years — reportedly in 88 hours.

The declaration is correct. It is also, in its current form, nearly powerless. And the reason has almost nothing to do with whether the mathematicians are right.

What Tao Is Actually Describing Is a Labor Dispute

He told New Scientist that mathematics has to clear five hurdles — a result must be created, checked, explained, accepted, and digested — and that the AI firms care only about the first two. His image for this was a communal village table piled with raw meat, the food problem declared solved by whoever dropped it and left.

That is not a philosophy-of-science complaint. That is a supply-chain complaint, and it is about who pays for the second shift.

A university mathematics department has always run on two currencies. Money pays for the teaching and the offices. Credit pays for everything that makes the field a field: the referee who spends three weekends hunting a gap in a proof, the seminar audience that finds it, the doctoral student who checks line 40, the survey article that compresses a decade into forty pages a graduate student can actually read. None of that is compensated, and it was never meant to be. You do the unpaid work because being the person who certified a result is worth more than money, and because someday someone certifies yours.

That system works precisely as long as the headline and the certification cannot be separated. The labs separated them. They take the headline on a Tuesday and leave the referee report on somebody’s tenure clock.

The obvious reply is that AI can check proofs too. Sometimes it can. But “verified by the same class of system that produced it” is not a certification regime. It is a mirror.

88 Hours Is Not the Number That Should Worry Anyone

The labs’ metric is hours-to-solve, because that is the metric they chose and the metric that makes a launch stream. Nobody publishes hours-to-digest. There is no benchmark for the years it takes a hard result to move from a preprint to something a competent graduate student can be taught, and there never will be, because you cannot demo a decade in a keynote.

Which is why the signatories had to call this an “alignment” problem. A generation ago they would have called it publication standards, or research misconduct, or simply the literature. Today the only vocabulary with moral weight in the room belongs to the AI safety people, so a declaration about refereeing arrives dressed in borrowed clothes. That tells you where the two fields now stand relative to each other, and it is not flattering to the mathematicians.

The Reflex the Right Should Resist

The instinct on my side of the aisle is to read any document signed by 25 medalists as a guild defending a moat, and to reach for the file marked rent-seeking. Usually that instinct earns its keep. Here it half-earns it.

What the labs are doing is the oldest move in the regulated-industry playbook: capture the value, externalize the cost. If you would not let a plant discharge into a river because it swears it is innovating, you should not let a lab discharge unverified theorems into the literature. The difference is that rivers have regulators. The literature has a website.

Watch for the response, and it will arrive as a press release: an “AI for science” program, compute credits, a partnership announcement with a university’s name in the second paragraph. That is not funding for verification. That is a retainer.

“There is a whole category of work that now has no owner,” a second-year postdoc at a large state university, who referees for two journals on her own time, put it in a Slack message to a co-author. “The lab gets the citation. I get the email at 11 p.m.”

What Would Actually Move the Labs

Mathematicians cannot fix this with a declaration, because the instrument that changes behavior is money and the signatories do not have any. Which means the realistic path is narrower and considerably less dignified than a manifesto: put a tariff on verification.

Journals and departments can decline to referee lab output on the community’s usual timetable. If a company wants its results digested, the company funds the digestion — endowed postdoc lines, named referee stipends, a budget line that pays for the checking rather than the computing. That is not a philosophical settlement. It is a rate card, and rate cards have the advantage of being enforceable by people who have never met each other.

The uncomfortable half of this, for the signatories, is that they have spent careers insisting mathematical knowledge is a public good maintained by a self-governing republic of letters. Fine. Public goods get underproduced the moment the toll is removed, and asking nicely is not a toll.

The labs will keep announcing. The medalists will keep declaring. Whether any of it changes comes down to a single question, and it is not a mathematical one: will anyone ever write a check for the checking?

Sources