Gowers: the maths problems models have solved were solved by counterexample, not proof

Fields Medallist Timothy Gowers draws a striking distinction on his blog: almost all the famous mathematics problems solved by language models to date were solved with counterexamples rather than proofs.

What the difference is

In mathematics, refuting a claim and proving one are two very different jobs.

Finding a counterexample means producing a single object that shows the claim is false. To demolish a claim that "every number with this property satisfies that", one number that fails is enough. This is a search task: you have to look in the right place, in the right way.

A proof means showing the claim holds across infinitely many cases. No single example will do; you must build a chain of reasoning covering every case. This is a construction task.

Why models are better at one

The distinction connects directly to how models work. Searching for a counterexample means moving quickly through a large possibility space, trying many candidates and discarding them. Language models are good at that: they are fast, they do not tire, and they try odd candidates a person would not think of.

Building a proof asks for something else — a structure in which every step follows necessarily from the last, with no gaps anywhere. A proof has no "probably" or "generally"; a single loose link invalidates the whole chain.

Why it matters

Gowers's observation shows how headlines reading "AI solved a maths problem" ought to be read. Two different achievements are described in the same sentence, and they are not the same thing:

• Counterexample — a real and valuable contribution. It can answer a question that has stood open for years, and the answer is verifiable.
• Proof — the actual work of mathematics. It explains why a theorem is true, not merely that it is.

This distinction is not a dismissal. Finding counterexamples is hard work and the models' success at it is real. But naming the contribution accurately is what makes the next step visible.

The verifiability question

The distinction has one more practical consequence. A counterexample is verifiable on its own: you take the example, check that it fails the claim, and you are done. How it was found does not matter.

A proof has to be checked line by line. A proof produced by a model may contain a hidden gap, and finding it takes the labour of a mathematician reading it. So a model producing proofs does not remove the verification burden — it merely moves it.

That also explains why models have advanced faster on the counterexample side of mathematics: there, the result verifies itself.

The same distinction elsewhere

The distinction Gowers draws in mathematics is a pattern with an equivalent in nearly every field where AI is used. Models are markedly more reliable at work whose answer is cheap to verify.

Whether a piece of code runs is settled by a test; whether a vulnerability is genuinely exploitable is settled by trying it; a hypothesis can be shown false by a single data point. In all of these the model produces and a separate mechanism verifies.

By contrast, in work where correctness can only be established by expert judgement — whether a strategy is sound, whether a text describes an event accurately, whether a proof is watertight — models are not advancing at the same rate. There, an error has no way of announcing itself.

The practical conclusion

The question to ask before handing work to AI is not whether the model can do it, but: can I verify the output cheaply? If the answer is yes, the model is a powerful tool. If not, the speed gained is repaid in verification labour.