In September 2026, OpenAI announced its models had cracked one of math's most famous problems. The proof checked out, but nobody understood it, and a field built entirely around solving problems has no system for that.
On September 8, 2026, OpenAI announced its models had solved one of the most famous problems in mathematics, and promptly set off an argument about authorship and standards. Scott Aaronson of the University of Texas, Austin wrote on his blog that human mathematicians are forevermore dethroned as the main theorem-proving entities on the planet.
Two days later, some 150 students, postdocs and professors filled a classroom at UC Berkeley. Ken Ono, who had left academia for an AI startup called Axiom Math, had come to talk about the shift. He told the room to brace. The room answered with anger. A 50-minute slot ran past two hours on questions alone. One student, hands trembling, asked him what he was actually doing.
Teleported to the summit, blind in the fog
What mathematics values is not the answer but the understanding. The author uses a mountain as the picture: an AI proof is like being teleported to the peak.
The proof is true. A system checks the logic. But you are standing in fog. You don't know where you are, how this mountain connects to any other, and you carry no tools to bring someone else up.
Climb it yourself and things are different. Your body adapts to thin air. You invent instruments. You get lost with a companion in a hidden valley and find a plant that becomes a medicine. That process is what mathematics actually produces.
The system was built for solving
For centuries, proving theorems and understanding them went hand in hand, so the whole culture of mathematics grew around problem-solving: graduate students cut their teeth on small problems, then publish, find jobs, earn recognition, all measured in theorems proved.
That incentive structure was not built for a world where solving gets easy and understanding stays hard. Tasmin Chu, a doctoral student at Caltech, puts it plainly: it takes our system to its absolute limit and destroys it.
The damage is already visible. OpenAI and Anthropic compete directly against mathematicians, and have hired many of them away, while keeping more than 100 important results secret, so nobody knows what was proved. The preprint server arXiv is flooded with AI-generated proofs, and has capped submissions at two per author per month. Some mathematicians no longer post open conjectures at the end of their papers, afraid the bots will scrape them.
Two futures
Daniel Litt of the University of Toronto describes the bad one: once the profession turns over, you get mathematicians trained to push a button without reading the output. If nothing changes, he says, there may simply be no more math in 50 years — not even done by machines.
The optimistic version hands the hard digging to AI and leaves humans to interpret, explain, build frameworks and ask new questions. Pavel Etingof of MIT says there will be far more mathematics to make sense of, and plenty of work for mathematicians, just different work.
The open question is whether the next generation still wants in. Many people choose pure math because they get to author their own creative contributions and hand them to a community that values them. If the daily reality becomes machines talking and humans listening, that pull disappears.
The author ends with hope: being forced to say out loud what mathematics is good for may leave the field stronger. Whether it gets there is up to the mathematicians.
Why it matters
If AI proves every possible theorem and no human understands any of them, it cannot be part of culture, because culture is inherently human. That moves the argument past whether AI replaces mathematicians to something harder: a correct answer nobody comprehends does not exist as far as civilisation is concerned.
Curated from high-quality sources, with concise summaries and key takeaways.
