Mathematicians just answered the machine — and it's the best model yet for what AI does to a profession

The framing that actually helps

Most commentary about AI and expert work runs on a single axis: will it replace people or not. Tao’s ICM talk sidesteps that question, and the sidestep is the contribution.

His argument, as reported from the congress: mathematics has always operated under scarcity of proof. Producing a genuine result is slow, difficult, and rare, and essentially every institution in the discipline — journals, credit, hiring, prizes — is built around allocating attention to a small number of hard-won results.

Machine-generated proofs do not break that system by being wrong. They break it by being abundant.

If proofs become cheap, the constraint inverts. The scarce resources become the ability to decide which results matter, to check that they are correct, to explain them to other humans, and to fold them into the canon that the next generation learns. Tao’s term for the new problem is proof overload, and his conclusion is that mathematicians will need new workflows and infrastructure to work with these systems on their own terms rather than on the vendors’.

That is a far more useful model than replacement, and it transfers. The generalisable claim is that AI does not eliminate a profession’s work — it makes one part of that work abundant and relocates the scarcity. Any knowledge worker can run the test: which part of my job is the bottleneck, and what happens to my role if that specific bottleneck goes away?

The historical comparison is deliberate

Tao compares the present to the foundational crisis of the early twentieth century — the period when paradoxes in set theory forced mathematics to rebuild its logical underpinnings, producing decades of anxiety and, eventually, a stronger discipline.

That analogy carries two implications, and both are worth taking. The optimistic one: the field has survived an epistemological earthquake before. The pessimistic one: it took roughly thirty years, and permanently changed what mathematicians did day to day.

Tao’s practical picture is a shift from writing proofs on blackboards toward writing code in Lean — mathematics adopting the tooling posture of software engineering. The productive arrangement is hybrid: models suggest paths, formal systems enforce truth, the loop self-corrects. Machine intuition is useless here without a mechanism that refuses plausible-sounding falsehoods, because mathematics’ correctness requirement is binary in a way almost no other field’s is.

What the profession actually said

The range of views reported from and around the congress is worth recording, because “mathematicians are worried” flattens something more interesting.

Working optimism. Abhishek Saha characterises frontier models as the equivalent of “a solid and indefatigable PhD student” — capable, tireless, requiring supervision — which lets a senior mathematician act more like a conductor than a soloist. That is a specific, operational description, and it matches what strong practitioners in other fields report.

Cultural worry. Timothy Gowers, a Fields medallist, has described genuine discomfort at seeing AI dispatch problems he had personally struggled with. His stated concern is not employment but the possible destruction of mathematical culture: if fewer humans ever develop deep expertise, the discipline could accumulate a vast literature that no living person actually understands.

That is the sharpest version of the risk, and it is not sentimental. A proof nobody comprehends is technically knowledge and practically inert. It cannot be extended, taught, or reused with judgement.

Existential objection. A critique circulating under the title “The Dark Night of Mathematics” argues the Leiden Declaration offers reassurance rather than confronting a deeper loss — that for many practitioners the point of mathematics was the encounter with something ineffable in creating it, which delegation forecloses. Whatever one makes of that, it is a real position held by real people and it does not reduce to Luddism.

The Leiden Declaration is the concrete part

Positions at a congress are one thing. The Leiden Declaration on Artificial Intelligence and Mathematics is the institutional response, and it is remarkably specific.

Published in June 2026, out of a September 2025 workshop at Leiden University’s Lorentz Center, with sixteen primary authors including Jim Portegies, Mateja Jamnik, Ursula Martin, Rodrigo Ochigame and Michael Harris. Over a thousand people signed on the first day; the count now exceeds 3,000.

It asks for three things:

1. Mandatory disclosure of AI use in research papers
2. Peer review for all AI-assisted mathematical work
3. Funding and legal support so academic researchers are not competitively disadvantaged relative to those with corporate-scale tool access

And it names concerns beyond correctness: that commercial interests could tilt research priorities toward problems that happen to be computationally tractable, that citation practices in trained models are inadequate, and that training-data collection has involved copyright violation. Michael Harris has framed the goal as recovering control of the narrative about mathematics’ values from the AI industry.

Note what is absent. There is no call to ban the tools, no demand that AI-assisted results be excluded, no claim that machine proofs are illegitimate. It is a demand for standards from a profession that intends to use the technology.

The disclosure split running through this week

Set the Leiden Declaration’s first demand next to what happened elsewhere in the same seven days.

Mathematics is moving to make AI disclosure mandatory in published papers — disclosure aimed at readers, enforced by the discipline.

Journalism went the other way. A record eight Pulitzer entries disclosed AI use to the prize committee, and many of those investigations carried no disclosure in the published stories at all. The judges knew; readers frequently did not.

Regulators split the difference. The EU’s Article 50 made reader-facing disclosure mandatory for AI-generated text on 2 August, while leaving research tooling untouched.

Three professions, one week, three different answers to the same question: who is owed an account of how AI was used? Mathematics says readers and reviewers. Journalism has so far said the awarding body. The EU says readers, but only when the machine wrote the words.

Mathematics is the only one of the three that arrived at its answer by argument rather than by regulation or by default — which is probably why it is the most coherent.

Why this matters

It is the clearest available template for AI in expert work. Not because mathematics is representative, but because it is ahead. The sequence — capability arrives, output becomes abundant, the profession discovers the bottleneck has moved, institutions scramble to define standards — is going to repeat in law, medicine, accounting and engineering. Watching it run once, in a field with unusually crisp success criteria, is worth more than a dozen think-pieces about the future of work.

Mathematics has an advantage nobody else gets. Lean can mechanically decide whether a proof is valid. That is why the discipline can absorb abundance without drowning — as the Astra results showed, the certificates make the output checkable by strangers. Every other profession faces the same flood without a verifier. When your field’s equivalent of proof overload arrives, there will be no Lean.

The access concern is the most underrated part. Leiden’s third demand is about funding and legal support, not correctness. If the models that produce research-level results cost more than a university department’s annual budget, the competitive landscape of a public discipline changes regardless of how good the tools are. That is the same structural worry visible in Google’s Atlas report on AI and the economy, arriving from the other end.

Professional bodies are becoming the real governance layer. Government frameworks are late and, in the US case, unpublished. Meanwhile a declaration written by sixteen academics has 3,000 signatures and concrete demands. As with the 1,100 AI employees who signed the slowdown letter, the fastest-moving norms are being set by practitioners, not regulators.

Honest caveats

Tao’s slides could not be read directly for this article — the published PDF is image-based and did not yield text. His positions here are drawn from reporting on the talk, not from the primary document, and the specific wording of his framework should be checked against the slides or a recording before being quoted precisely.

“3,000+ signatories” is a reported figure and signature counts on open declarations are weak evidence of anything beyond attention. Mathematics has roughly a hundred thousand active researchers worldwide.

The reported views are a sample, not a survey. Saha, Gowers and the “Dark Night” critique are individual positions selected by journalists for contrast. No systematic poll of mathematicians’ attitudes exists, and the quiet majority is unrepresented in every account of this, including this one.

Nothing here is settled. Whether AI-generated results genuinely become abundant, or whether the current run of solved problems reflects low-hanging fruit, is unknown. The hardest problem categories and the six unsolved Millennium Prize Problems remain untouched, which is at least consistent with the low-hanging-fruit reading.

What to take from it

If you do knowledge work: run Tao’s test on your own job. Identify the step that is currently the bottleneck, then ask what your role becomes if that step gets cheap. The answer is rarely “unemployed” and rarely “unchanged.”

If your field has a verifier, invest in it now. Compilers, type systems, test suites, reconciliation processes, formal review. The reason mathematics is handling this comparatively well is that it can mechanically reject a wrong answer. Anything that gives your field a version of that gets more valuable as output becomes cheaper — a point that also runs through why vendor benchmarks keep failing.

If you are setting policy for a team using AI: the Leiden template is a good starting point and it is free. Disclose use, review the output, and make sure access is not concentrated in whoever has the biggest budget. Those three rules would survive translation into almost any professional context.

Related: Astra’s ten proofs and the verification lesson · A record eight Pulitzer entries disclosed AI · July 2026 in AI — what changed for buyers

Questions or corrections? Email Pick Right. Want the full list? See all news.