An AI Just Disproved an 80-Year-Old Math Conjecture

For roughly 80 years, mathematicians assumed a particular conjecture was true. On May 20, OpenAI announced that one of its models had found a counterexample. This isn’t another “AI solves a hard problem” headline. It’s closer to the first time an AI has overturned established mathematical consensus.

Not solved — disproved

The conjecture sits in discrete geometry, part of the sprawling list of open problems left behind by Paul Erdős, the Hungarian mathematician who posed more questions in the 20th century than almost anyone else. This particular one had been quietly assumed to be true since the late 1940s.

The distinction matters. Proving a conjecture and disproving one are different games. Disproof needs exactly one counterexample — but finding that one example is what stumped the field for eight decades. The model worked through structural territory human mathematicians had never seriously explored, and constructed a counterexample there.

From solving to discovering

Until now, AI in mathematics has lived in two boxes. One is competition-style problem solving — IMO-grade questions with known answers. The other is proof assistance, where the AI plays helper to a human-driven argument. Both are humans posing the question and largely defining the path.

This is different. The model designed its own search space and built a counterexample in a region no human had thought to probe. OpenAI’s accompanying video, “The Erdős Breakthrough,” cleared 4,000 views and 440 likes within a day — drawing attention from both the math and AI sides of X and Hacker News simultaneously, which rarely happens.

Why discrete geometry first

Discrete geometry studies the arrangements of discrete objects — points, lines, polygons. It’s visually intuitive but combinatorially brutal. Cases explode the moment you try to generalize. For human mathematicians, it’s the classic “I can see it, but I can’t check every configuration by hand” zone.

That’s exactly where current AI shines. Pattern-finding across enormous combinatorial spaces, probing the corners human intuition can’t reach. Discrete geometry may well be the first branch of mathematics where AI overtakes humans at the frontier.

Are mathematicians out of a job?

Short answer: no, and not soon. A counterexample is a starting point, not an endpoint. Understanding why it works, generalizing it, and weaving it into broader theory — that’s still human labor. The real signal here is different.

The pace of mathematical discovery can now compress from a human lifetime to a single inference run. Eighty years collapsed into days. The mathematician of the next decade will likely spend their time interpreting what the AI surfaces and building theory around it, not racing the machine to a proof.

The lingering question

“AI does math” isn’t metaphor anymore. It’s a tool that overturns consensus and surfaces new facts. Which conjecture falls next? And the harder question — when AI finds a mathematical truth, will humans still be able to fully understand what it found? An 80-year-old assumption just collapsed. The next 80 years of mathematics start in the rubble.