An 80-year-old math problem. The world's best mathematicians had no solution. Then an AI came along — and solved it.
No joke. No PR stunt. An OpenAI model disproved a conjecture by Hungarian mathematician Paul Erdős on May 20, 2026 — a problem that had been open since 1946. The proof was reviewed and confirmed by external mathematicians.
What is the Erdős problem?
Paul Erdős was one of the most prolific mathematicians of the 20th century — famous for posing hundreds of tricky problems and leaving them for others to solve. His so-called unit distance problem sounds simple:
Imagine n points scattered on a sheet of paper. How many pairs of those points can be at exactly the same distance from each other?
Erdős conjectured that a square grid — think of graph paper — is nearly optimal for this. No other arrangement can produce significantly more equal-distance pairs.
Sounds like a neat puzzle. It went unsolved for 80 years.
What did the AI do?
OpenAI's reasoning model disproved Erdős' conjecture. It found an arrangement that outperforms the square grid — meaning the conjecture was wrong.
What's remarkable isn't just the result, but how it got there. The model didn't run a brute-force search. It applied deep algebraic number theory — a field that even many professional mathematicians only know in outline. The proof was independently verified by Princeton mathematician Will Sawin and confirmed as correct.
The model didn't assist. It solved.
Why does this matter?
Because it's the first time an AI has autonomously solved a significant open math problem.
AI and mathematics — that's not new. But until now, the pattern was always: human has an idea, AI checks it, structures it, helps formulate it. This was different. The model found an approach that mathematicians had missed for 80 years.
This isn't an assistance win. It's a discovery win.
The honest take
Time for the sober part — AI breakthrough headlines always need a reality check.
The Erdős problem has a property that plays to AI's strengths: it's precisely formulated, has clear rules, and right or wrong can be checked mechanically. That's exactly the type of problem where AI excels.
Do many open math problems look like this? No. The most famous ones — the seven Millennium Prize Problems from the Clay Mathematics Institute, each worth a million dollars — are far less structured. Whether AI can help there is completely open.
What the AI did not do here: decide which problems are worth solving. Explain why the result is significant. Put it in context for what to search for next. That remains human work.
In short: AI just won an 80-year sprint. The marathon — intuition, meaning, context — we're still running ourselves.
What does this mean for the rest of us?
Directly: not much yet. This is basic research, not a product.
Indirectly: a lot. If AI starts making autonomous discoveries in science — not just assisting — it changes how research works. Faster, broader, in areas where human intuition hits its limits.
OpenAI has announced it will deploy the model on more open problems. The math community is both excited and skeptical — a healthy combination.
One thing is definitively false since May 20, 2026: "AI can only generate text."