AI is solving 80-year-old math puzzles, and some mathematicians feel 'the rug pulled out'

The math world just hit a tipping point. In May 2026, OpenAI published a counterexample to the Unit Distance Conjecture, a geometric graph theory problem that had stood open since 1946 and was linked to the legendary Paul Erdos. That proof didn’t just collect dust — within a week, human researchers grabbed the core technique and used it to kill another major conjecture. Since then, the floodgates have opened. Epoch AI’s brutal FrontierMath benchmark recently logged its second solution to an open problem, and OpenAI’s new Astra model dropped with ten solutions of varying difficulty.

Abhishek Saha, a math professor at Queen Mary University of London, describes frontier models as being at least as good as a tireless PhD student. He burned a full day with GPT-5.5 Pro doing work that would’ve taken him weeks, leaving him feeling less like a solo performer and more like a conductor. But that orchestra analogy cuts both ways. Fields Medalist Timothy Gowers predicted this exact golden age back in 2000, where computers handle the grunt work and humans focus on deep ideas. He also predicted it wouldn’t last. He now writes about the deeply unsettling experience of having GPT 5.6 Pro solve a problem on its first try after he’d sunk serious time into it, calling it “not particularly pleasant to have the rug pulled out from under my feet.”

What really scares Gowers isn’t being out-computed — it’s the potential collapse of mathematical culture. If AI churns out a vast literature that fewer and fewer humans can actually understand, the whole enterprise becomes hollow. He’s wrestling with this publicly around the Leiden Declaration on AI and Mathematics, a transparency-and-ethics push signed by over 3,000 mathematicians and backed by the International Mathematical Union. Gowers supports the spirit of the thing but hasn’t signed it himself.

Trefor Bazett, writing in The Conversation, points to an April 2026 Carnegie Mellon paper that cracked a Ramsey theory problem by chaining together SAT solvers, LLM-generated code, and formal proof verification — the exact toolkit Gowers once imagined. This is genuinely new. But Bazett also notes that AI’s aptitude is wildly uneven across subfields; it’s a beast in graph theory and stumbles elsewhere. The hardest problems are still out of reach, which means the real story here isn’t AI taking over math. It’s that a profession built on human insight now has to negotiate a working relationship with a partner that doesn’t think like a human at all.