Gowers breaks with Fields medallists' AI warning, but agrees math faces a crisis

Mathematician Tim Gowers explains why he declined to sign an open letter, endorsed by 25 Fields medallists, that warns AI could harm mathematics. The letter’s core claim is that problem-solving is merely a proxy for the real goal—conceptual understanding and insight—and that machines churning out true/false statements at ever-greater speed could destroy the fertile ground where new ideas grow. Gowers agrees a crisis is underway but rejects this framing. Drawing on his decades-old essay ‘The Two Cultures of Mathematics,’ he argues mathematicians sit on a spectrum between problem-solvers and theory-builders, and that the letter risks implying only the conceptual-understanding camp holds the ‘right’ temperament—effectively telling a large share of the field, himself included, that their instincts are wrong.

Gowers is unusually direct about potential conflicts of interest, a notable move given how AI debates often collapse into accusations of bias. He discloses that he has contacts in OpenAI’s math group, gets early access to unreleased models and free Pro accounts, but has never been paid by the company. He also runs a Cambridge automatic-theorem-proving group—which he frames as a reason to be skeptical of AI, not aligned with it, since LLMs can now prove theorems without the human-insight-driven methods his team spent years developing. That, he says, is the ‘bitter lesson’ his group has had to swallow, and it arrived faster than anyone expected.

Rather than join an opposing camp, Gowers frames his post as an attempt to avoid one. His stated worry is that the mathematical community could fracture into bitter factions over AI, and he wants to reframe the disagreement as a difference in analysis rather than values. The deeper question he poses—whether it matters more that collective mathematical understanding advance as far as possible or that as many problems as possible get answered—cuts to how the profession should measure its own worth once machines can generate proofs on their own.