Did an AI Finally Solve a Math Conjecture That Stumped Mathematicians for 80 Years?

Yes — an AI produced a counterexample to a geometry conjecture proposed by mathematician Pál Erdős in 1946, disproving a problem that had remained unsettled for eighty years and carried an unclaimed $500 prize.

The Conjecture That Defeated Human Mathematicians

In 1946, Pál Erdős — one of the most prolific mathematicians in history, with over 1,500 published papers to his name — proposed a geometry conjecture that seemed deceptively simple. For decades, it sat unresolved. The world’s best mathematical minds attempted it and failed. The conjecture even had a $500 prize attached to it, offered by Erdős himself, to anyone who could settle it. That prize went unclaimed for eighty years.

Then an AI cracked it.

What Is the Erdős Conjecture and Why Does It Matter?

The conjecture in question belongs to the field of combinatorial geometry, an area concerned with how geometric objects — points, lines, distances — can be arranged and counted. Erdős had a gift for posing problems that were easy to state but extraordinarily hard to resolve. Many of his conjectures remain open today. The fact that this one resisted solution for eight decades is not unusual for Erdős problems — but what finally broke it is.

An OpenAI model produced an explicit counterexample: a concrete configuration that demonstrated the conjecture was not universally true. In mathematics, a single valid counterexample is all it takes to disprove a conjecture entirely. The result wasn’t a proof that the conjecture was true — it was a demonstration that it was false, which is in many ways a more decisive outcome.

How AI Reached This Level of Mathematical Reasoning

The breakthrough didn’t arrive in isolation. It followed a rapid acceleration in AI mathematical capability, most visibly demonstrated by Google DeepMind’s AlphaGeometry system, published in the journal Nature in January 2024.

AlphaGeometry trained on 100 million synthetic geometry problems — every single one generated without human-labeled examples. The system taught itself geometry from scratch. When tested on International Mathematical Olympiad (IMO) geometry problems, it solved 25 out of 30, approaching the performance level of human gold medalists. The previous best automated system had solved only 10.

That benchmark result signaled something significant: AI was no longer just a tool for verifying human mathematics. It was beginning to navigate open mathematical territory independently.

A Counterexample Is Not the Same as a Proof

It’s worth being precise about what happened here. Disproving a conjecture by counterexample is mathematically rigorous and definitive — but it’s different from constructing a full proof of a new theorem. The AI didn’t write a narrative explanation or develop a novel theoretical framework. It found a specific case that broke the rule Erdős proposed.

That distinction matters, but it doesn’t diminish the result. Finding counterexamples to long-standing conjectures requires searching an enormous space of possibilities. Humans had searched that space for eighty years. The AI found what they missed.

What This Means for the Future of Mathematics

Mathematicians were genuinely stunned. Not because AI solving problems was surprising in principle, but because of the target: a conjecture from one of mathematics’ most celebrated figures, one that had resisted the field’s best efforts across eight decades.

The implications are significant. AI systems are now capable of contributing original results to mathematics — not just checking proofs or performing computation, but navigating unsolved problems and finding answers humans couldn’t. This raises serious questions about the future role of human mathematicians, the pace at which open problems might now be resolved, and how credit and discovery are defined when the solver is a machine.

Eighty years. Fifteen hundred papers. One AI. The answer had been out there all along — it just took a different kind of mind to find it.