Friday’s most-argued-over document in mathematics was not a proof. It was a complaint about bookkeeping. “A Severe Misalignment of AI in Mathematics,” posted at mathandai.org, drew 689 points and 719 comments on Hacker News by Friday evening — and a signatory list of Fields Medal recipients that two same-day accounts could not agree on: 24, per a widely shared post from the Cornell mathematician Steven Strogatz; 25, per the AI Governance Institute’s write-up. When two dozen laureates can’t settle a headcount, you are watching a profession, not a movement.

And the profession’s real complaint isn’t that AI is bad at math. It’s that AI is good enough to break the accounting.

The Bottleneck Was Never Solving

Terence Tao has laid out the life of a proof in five stages: pose the problem, solve it, verify the solution, communicate it clearly, then get the community to digest and accept it. Models are now very fast at one and two. Stages three through five are done by people, in small numbers, with day jobs, and there is no capacity plan for a world where steps one and two get a thousand times cheaper.

Compute scales. Referee pools don’t.

“The department has exactly one person who can check this subfield, and last spring I learned it was me,” a postdoc at a large public university wrote in a Slack message at 1:40 a.m., in a thread that began with a preprint nobody could place. “The queue is triple what it was. There is no second referee.”

That is the misalignment, and it is not a philosophy problem. It is a staffing problem dressed as one — and everyone with a benchmark to sell has an interest in keeping it dressed.

Attribution Is the Guild’s Currency, Not the Public’s

The second grievance is attribution: unreferenced machine-generated arguments eroding the norm that a result belongs to a person. Fair enough — but note who is injured. Citation credit is the internal currency of a profession whose external customers are cryptographers, engineers, and the occasional physicist who needs a number she can stake a bridge on. Those customers have never cared whose name is on the theorem. They care whether it’s true, and they have no way to check.

This is where the signatories accidentally make the strongest market argument in the document — against themselves. One account of the earlier Leiden Declaration on Artificial Intelligence and Mathematics put its signatory list above 2,300; another reported 150 professors from Europe, Japan, and the U.S. Its five-item threat list leads with “plausible but unreliable outputs” and then moves to attribution, which is a diplomatic ordering. The unreliable outputs are the emergency. Attribution is the payroll.

And the guild built that payroll on volume. Academic mathematics has been measured for four decades by papers published, not errors caught. Flood the system with cheap plausible proofs and the metric doesn’t break — it inflates. The people with the most to lose from unreferenced AI arguments are also the people whose promotions, grants, and named chairs are denominated in reference counts. AI didn’t corrupt that incentive. It industrialized it.

Send the Bill to Whoever Is Wrong

Here is the part that should make everyone uncomfortable, including the people cheering the disruptors.

Verification is a cost. Someone has to bear it, and right now nobody does. Grants pay for novelty, never for checking — the money that funds a paper does not fund the referee who catches its error. Journals lean on unpaid labor and call it peer review. Industry labs publish results whose audit trail stops at a blog post and a demo. Every party in the chain has an incentive to move fast and let the next person find out.

Declarations don’t fix that. Prices do.

If funders carried verification as a budget line — not as service, as line item — queues get shorter and referees get paid. If journals charged for the audit and put their institutional names behind it, buyers would quickly learn which venues are worth the fee. If the labs that want their mathematics taken seriously bought the checking instead of the announcement, the 719-comment threads would get shorter and duller, which is the goal. And if a wrong result carried a named, accountable party, the market would do what it always does with unpriced risk: discover the price, painfully.

The signatories are right that something is misaligned. They just named the wrong party. The gap isn’t between AI and mathematics. It’s between who collects the credit and who absorbs the cost of being wrong — and the second group has been working for free since long before anyone trained a model on a proof.

Sources