Beauty Is Not a Warrant

Both were prompted by the same anxiety. Both produced answers of a strikingly consistent kind.

Venkatesh reframed the discipline away from proof-production: mathematics is "more akin to storytelling," he said, and "perhaps the proofs are part of the grammar." Vakil, asked from the floor whether aesthetic judgement could be mechanized, declined: "We know what's aesthetically pleasing. I don't think that's mechanizable." Kontorovich, pressed on whether a machine could be said to understand anything, refused the frame entirely — "I don't know what the word understands means." Demers, explaining why she will not let a model draft a difficult email, said the difficulty is the point: writing it is how you engage with what makes it hard.

Every one of those answers points at an interior. Understanding. Story. Beauty. The felt work of struggling with something until it gives. I take them to be honest answers — accurate descriptions of what these people actually experience when they work. But they are all reports from inside a head, and that is worth sitting with, because a warrant that depends on interiors is precisely what mathematics spent two millennia arranging not to need.

The warrant they are walking away from

The reason mathematics occupies the position it does — the most trusted form of knowledge humans produce, the thing other disciplines reach for when they want to sound certain — is not that mathematicians are unusually careful people. It is that mathematics made its checking external.

A proof is a proof to someone who does not share your taste, does not like you, does not care what you were trying to say, and cannot follow your intuition. That was a deliberate construction, built over centuries and pushed hardest in the formalist program: reduce the thing to symbols and rules for moving them, so that verification requires nothing of the verifier except patience. Whatever else it cost, it bought a criterion that settles arguments between strangers.

The field has tested that criterion against exactly this pressure before, and it held. When Kenneth Appel and Wolfgang Haken proved the four-color theorem in 1976, they did it by reducing the problem to nearly two thousand configurations and having a computer grind through them. No human read the whole proof. No human ever will. It was widely described as ugly, and the objection that it produced no understanding was made loudly and by serious people. The theorem was accepted anyway, because the criterion said accept. Beauty did not get a vote.

The obvious rebuttal is that nobody on that stage was voting on a theorem. Strogatz asked what mathematics is for, which is a question about value, not validity — and on validity the panel conceded nothing. Every one of them would accept a Lean-verified proof from a machine they found hideous. Vakil's claim that aesthetic judgement cannot be mechanized is a claim about which problems are worth attacking, not about which arguments hold.

That rebuttal is correct. It is also the trap. The two axes are genuinely distinct, and they stay distinct right up until somebody with budget authority asks a discipline to justify itself — at which point they collapse, because a funder is not evaluating theorems. Validity is what mathematics can prove to a stranger. Value is what it has to prove to a stranger in order to get paid. And the field's whole two-millennia advantage sits on the axis nobody is attacking.

Which is why the timing matters, and why Vakil identifies the real threat correctly and then hands the wrong tool to meet it. He names two worries: students who route around the struggle, and anti-intellectualism using AI as "a Trojan horse" for defunding science. The second is the serious one. It is also the one where an interior is least useful. You cannot fund a discipline on a quality of experience available only to its practitioners. The person holding the budget has no access to it, and the person trying to cut the budget has an obvious rejoinder ready: if it is only meaningful to the people doing it, let them pay for it themselves.

That is the shape of the problem. Not that mathematicians have stopped believing in proof — they plainly have not — but that the public account of what the discipline is worth has quietly relocated from the one place it was checkable to the one place it is not, at the exact moment somebody started asking.

The calculator is not the analogy

"The best person with a calculator is the best person without a calculator," Vakil said, and it is the most reassuring sentence in either conversation. It is also an empirical claim wearing the clothes of a principle.

The calculator sits downstream of judgement. You decide what to compute; it computes. It has no opinion about which computation is worth doing, and it never proposes one. Under that arrangement the amplification claim is close to trivially true: a device that multiplies whatever you bring will preserve the ranking of what people bring.

Kontorovich breaks the analogy in the same session, describing what current systems are actually good for. Feed a model a thousand problems, he says, and you might get ten good papers a year out of the pile. That is not a device executing a specified operation. That is a collaborator with a search strategy, and the strategy determines which problems come back looking tractable, which determines which get attempted, which determines what the literature is eventually made of. The tool is upstream. It is shaping the question set before judgement gets a look.

Once a tool sits upstream of judgement, "it amplifies the best" stops being obviously true and becomes something you would have to go and measure. It might amplify whoever is best at operating it, which is a different ranking. Or it might advantage whoever is quickest at recognizing a good result they did not produce — a real skill, genuinely valuable, and not the same skill as producing one. The discipline has no established way to credit it, and its entire status system is built around the other thing.

Kontorovich is probably right that the machine lacks what he calls the human valuation mechanism — the sense of which results matter. But that mechanism was calibrated on a supply of problems that humans generated, under human constraints about what looked worth a year of somebody's life. If the supply changes composition, the calibration is being asked to do a job it was not fitted for. Nobody in the room claimed otherwise. Nobody was asked.

The physicist's answer is the only load-bearing one

Demers gives the answer that survives contact with the world, and she gives it without touching the metaphysics at all.

She does not argue about understanding. She says: validate the output, and "a physicist is responsible for what is in their publication if they're going to sign their name." That is procedural. It is a liability rule. It requires no one to settle what a machine knows, because it attaches the consequence to a person and leaves the question open indefinitely. It works whether the model is a stochastic parrot or a nascent mind, and it keeps working as the capability moves.

Note the inversion. The physicist supplies the operational answer; the mathematicians supply the aesthetic one — and mathematics is the field where the operational answer is most available. Lean, Mathlib, the whole formal-verification effort that Terence Tao has spent years pushing into the mainstream of the discipline: that infrastructure exists specifically to make "did the machine's proof actually check out" a mechanical question with a yes-or-no answer. Mathematics is the one field that can, in principle, verify machine output completely. It answered with beauty anyway.

Which suggests the panel is not really worried about correctness. Correctness is the solved part, or at least the tractable part. The worry underneath is about what happens to a life spent doing something once the doing stops being scarce — and that is a much harder thing to say out loud in a convention center than "we have concerns about reliability." So it comes out as storytelling and aesthetics, which are true, and which are also the available vocabulary for a grief that does not have its own words yet.

Demers's other position is not color, and should not be handled as color. She has never used Claude or ChatGPT. She also calls what happened to the training corpus "outrageous" and names it a moral hazard. Those are not two facts; they are one position. She is not declining on capability grounds, she is declining on provenance grounds, and that is a coherent stance rarely stated plainly by someone in her seat. It does mean her account of what these systems cannot do is testimony from outside the room. That weakens the empirical half and leaves the ethical half entirely intact — the moral claim does not depend on the capability claim and should not be read as standing or falling with it.

What the corpus is made of

Here is what actually connects the two conversations, and it is not the anxiety.

The systems in question are made of mathematics. Not a theory of mathematics — the accumulated output itself: the textbooks, the arXiv preprints, the lecture notes, the patient MathOverflow answers written by people explaining something to a stranger for nothing. Venkatesh says the discipline's real work is storytelling and the conveying of understanding. That exposition is exactly what got pooled. The corpus these models are built from is the discipline's most human, least mechanical layer, flattened into weights.

So the field is worried about a machine that is a compression of the field. When the panel asks whether it understands, they are asking whether a summary of everything they collectively wrote understands — which is a strange question, and probably not the one that pays.

The better question is the one Demers already asked from a different direction: who was consulted.

The mathematical inheritance is genuinely a commons. It was built largely by people who gave it away, and giving it away is the tradition rather than an accident — nobody wrote a MathOverflow answer for the money. That habit of open exposition is precisely what made the corpus available to be taken. The discipline's gift economy is the reason the training data existed, and the discipline is now in the position of having that gift sold back to it as a subscription product by parties who contributed none of it.

That is not a story about understanding. It is a story about enclosure, and unlike the other one it comes with mechanisms attached: attribution, licensing, compensation, collective bargaining over a corpus a profession collectively produced.

The one-way valve

Except that every mechanism on that list works the same way. Each converts a freely given commons into property that can be withheld, because withholding is the only leverage property has. A commons defended by licensing is not a commons afterward; it is a rights pool with better manners. Which is the actual bind, and it is worth saying rather than stepping around: the gift economy has no defensive weapon that does not destroy the gift. Mathematicians cannot fight enclosure with enclosure and still be the people who wrote those answers for nothing.

So the harm needs describing more precisely than "they took it." A commons is supposed to be usable. Being drawn from is what it is for, and a discipline that starts objecting to use has abandoned the tradition faster than any lab could take it from them. The harm is that the taking ran one way. The input stayed open and the output closed — two million lines of Mathlib, decades of exposition, the whole gift in one end; a subscription at the other. It is not the use that breaks the arrangement. It is the valve.

That reframes what there is to ask for. Not payment, which concedes the property frame and requires mathematicians to become the sort of people who charge. Openness: if you are built out of a commons, you return to the commons. And the demand is not hypothetical, because the line is already being drawn inside the discipline by people who did not wait for a panel to authorize it. Mathlib is open, and the labs are on both sides of it — DeepMind trained AlphaProof on formalized mathematics in Lean; Harmonic made an inaugural donation to the Lean FRO in February; Amazon Web Services announced the largest donation in that organization's history in July. Gift taken and gift returned, through the same infrastructure, in public, with the terms legible to anyone who wants to check them.

What mathematics is for can be answered many ways and settled none. Whether the things built out of a discipline's donated work owe anything back to it is a question that can actually be decided, by people, in rooms, this decade — and is being decided right now, mostly by default. The panel spent an hour on the first. Demers, who has never opened the software, spent a sentence on the second. Between them, she is the only one who named something anybody could go and do.