Did an AI Just Make Mathematicians Obsolete?

The news that Claude produced a machine-checked proof of Fermat's Last Theorem travelled fast, and by the time it reached social media the story had grown a few sizes. AI has solved mathematics, ran one version. Human mathematicians are finished, ran another. Both are wrong, and it is worth being precise about why, because the actual achievement is impressive enough without the inflation.

What the AI did, and did not do

Start with the thing itself. Claude did not discover a new proof of Fermat's Last Theorem. It took the proof that Andrew Wiles published in 1994, a proof mathematicians already accept, and translated a version of it into Lean, a language that forces every logical step to be spelled out so a computer can check it. That job is called formalization, and it is the mathematical equivalent of turning a finished building's blueprints into a structure verified brick by brick. Valuable work. Not the same as designing the building.

The distinction matters because the hard, creative part of mathematics is coming up with the argument in the first place. Wiles spent seven years on that, mostly in secret. What Claude compressed from years into days was the translation, not the insight. Anthropic itself is clear that the model followed a simplified route through an existing proof, leaned on a large library of already-formalized mathematics, and took occasional human guidance when its agents got stuck.

Where the fear comes from

The worry that machines will replace mathematicians is old, and it usually rests on a confusion between two things a mathematician does. One is verification, the careful checking that every step holds. The other is discovery, the leap to a new idea worth checking. Computers have been chipping away at the first for decades, and formalization tools are the latest and most powerful example. There is little sign yet that they can do the second on their own.

This is the same category error behind the recurring question of whether a model that predicts text can be said to actually understand anything. Producing valid mathematical steps, even a great many of them very quickly, is not the same as knowing which questions are worth asking. Kevin Buzzard, the mathematician who reviewed the proof, framed the result as a tool for mathematicians rather than a replacement for them: a way to build a verified foundation the field can trust and search.

The honest version is still a big deal

None of this is a reason to shrug. The bottleneck in formal mathematics has long been human labour, the sheer tedium of writing out proofs in a form a computer will accept. If models can now clear that bottleneck at speed, large parts of the mathematical literature could be formalized and checked, which is genuinely useful. That is a real change to how mathematics gets done. It is just not the same as a machine sitting down and out-thinking the people who do it.

So the claim to be skeptical of is not that AI helped with a famous proof. It plainly did. The claim to reject is the leap from there to a world without mathematicians. On the evidence so far, the machines are getting very good at checking the work. Deciding what work is worth doing is still ours.