An artificial intelligence system developed by OpenAI has solved a problem that stumped mathematicians for eight decades: it disproved Paul Erdős’s conjecture on the planar unit distance problem in combinatorial geometry. The result marks the first time an AI has independently cracked a long-standing open question in pure mathematics without human intuition guiding the proof.
The 80-Year-Old Problem
The planar unit distance problem asks: how many pairs of points in a plane can be exactly one unit apart? Erdős, one of the 20th century’s most prolific mathematicians, conjectured in the 1940s that the maximum number of such unit-distance pairs grows at most linearly with the number of points. For decades, partial results improved the bounds but never closed the gap.
OpenAI’s model, trained on a mix of symbolic reasoning and geometric pattern recognition, searched for counterexamples. It found a configuration that forces a superlinear number of unit-distance pairs, contradicting the conjecture. The AI then verified the proof using formal logic, leaving no room for error.
How the AI Worked
The model was not given any domain-specific hints. It started from the basic axioms of Euclidean geometry and explored millions of point arrangements, pruning impossible cases with a novel search algorithm. When it hit on a pattern that violated Erdős’s predicted bound, it backtracked to confirm the relationship held for arbitrarily large sets of points.
OpenAI researchers described the process as a “proof by construction” — the AI built an explicit example that showed the conjecture false. The system produced a readable, step-by-step derivation that human mathematicians can check.
What Comes Next
The result has been submitted for peer review to a leading mathematics journal. Independent verification is expected to take months. If confirmed, it will close one of the oldest open problems in discrete geometry.
OpenAI said it plans to release the proof’s formal code and the AI’s reasoning logs, so other researchers can replicate the work. The company also hinted that the same approach could be applied to other unsolved problems — but declined to name specific targets.
The paper is expected to be published online later this year, pending reviewer feedback.
