AI Models Disprove 80-Year-Old Erdős Unit Distance Conjecture

For nearly eight decades, discrete geometers operated under a shared assumption about how points behave on a flat surface. If you scatter a set of dots on a piece of paper and connect every pair that is exactly one unit apart, the resulting web of lines was believed to grow at a very specific, slow rate. This was the Erdős unit distance conjecture, proposed in 1946, and generations of mathematicians assumed the optimal arrangements would always look roughly like square grids.[8][9]

That consensus collapsed in May 2026. An internal reasoning model developed by OpenAI autonomously generated a counterexample that definitively disproved the 80-year-old conjecture. By drawing on high-powered algebraic number theory rather than traditional geometric grids, the AI constructed a configuration of points that yielded significantly more unit distances than Paul Erdős had predicted possible.[1][8]

To understand what the AI accomplished, one must understand the mechanics of the problem itself. The planar unit distance problem asks a deceptively simple question: given a specific number of points in a two-dimensional plane, what is the maximum possible number of pairs that are exactly one unit apart?[2][7]

For small numbers of points, the answers are trivial—three points can form an equilateral triangle, yielding three unit distances. But as the number of points scales toward infinity, the complexity explodes. Erdős originally proved a lower bound by using a square grid, showing that a set of points could have a specific density of unit distances. He conjectured that this was essentially the ceiling, formally stating that the maximum number of pairs would grow just slightly faster than linearly.[1][2]

The mathematical community struggled to close the gap between Erdős's lower bound and the absolute upper limit. In 1984, mathematicians Joel Spencer, Endre Szemerédi, and William Trotter established a definitive upper bound. For forty years, human efforts largely focused on trying to lower that ceiling to match Erdős's prediction, rather than searching for a counterexample that would shatter it.[2]

The OpenAI model succeeded precisely because it was not constrained by the historical bias toward grid-like structures. Given the problem by researchers, the model engaged in a long chain of reasoning that bridged discrete geometry with algebraic number theory. It discovered an entirely new family of algebraic constructions that outperformed the traditional grid models.[1]

The AI's output was a several-page mathematical argument constructing a set of points with a growth rate that included a tiny positive constant, approximately ten to the power of negative thirty-eight. Because this growth rate strictly exceeds Erdős's predicted ceiling, it served as a definitive counterexample, proving the original conjecture false.

The discovery immediately shifted the role of human mathematicians from generators to verifiers. Fields Medalist Tim Gowers and other experts reviewed the AI's proof and confirmed its validity. Crucially, the AI's breakthrough served as a foundation for further human insight. Shortly after the AI's result was verified, Princeton mathematician Will Sawin improved upon the model's construction, pushing the lower bound up significantly higher.[6][11]

The OpenAI result was not an isolated anomaly. Within days, Anthropic reported that its unreleased Claude Mythos model had independently solved the same problem over a single weekend. Anthropic engineers ran isolated instances of the model with internet access blocked to ensure it was not simply retrieving OpenAI's leaked solution, confirming that frontier models now possess the intrinsic reasoning capacity to crack the problem.[5]

The same week saw a parallel breakthrough from Google DeepMind. DeepMind's AlphaProof Nexus system, which pairs a large language model with the Lean formal proof assistant, autonomously solved nine other open Erdős problems. Unlike the natural-language proofs generated by OpenAI and Anthropic, DeepMind's system produced machine-checkable formal proofs, eliminating the need for human verification.[5][10]

The evidence supporting these breakthroughs is robust, anchored by the public verification of the proofs by leading combinatorialists. Thomas Bloom, who maintains the official registry of Erdős problems, validated the findings, noting that the AI succeeded by persevering down complex algebraic paths that human researchers had previously dismissed as unpromising.[8][10]

However, the AI did not solve the unit distance problem in its entirety. While the models successfully disproved Erdős's lower bound, they did not establish a new, tight upper bound. The absolute maximum number of unit distances remains unknown, and the 1984 upper bound still stands. The AI showed that the floor is higher than expected, but it did not find the ceiling.[8][11]

The resolution of the unit distance conjecture marks a structural shift in how mathematical research is conducted. As AI models demonstrate the ability to cross-pollinate ideas between disparate fields—like geometry and number theory—the bottleneck in mathematics is moving from the generation of novel ideas to the rigorous verification and expansion of machine-generated proofs.[4][11]