AI Disproves 80-Year-Old Math Conjecture by Connecting Unrelated Fields

For eight decades, a deceptively simple geometry question posed by the legendary Hungarian mathematician Paul Erdős stood as an immovable ceiling in discrete mathematics. Known as the planar unit distance problem, Erdős's 1946 conjecture asked a fundamental question: if you place a certain number of dots on a flat piece of paper, what is the maximum possible number of pairs that can be exactly one inch apart? Generations of the world's sharpest mathematical minds attempted to crack the problem, broadly accepting Erdős's educated hunch about its upper limits. But in late May 2026, the mathematical community was jolted by a 125-page proof that shattered that 80-year-old assumption, proving that the accepted ceiling was fundamentally incorrect.[1][2][5]

The disproof did not come from a tenured professor at a prestigious university, nor did it emerge from a specialized mathematical software program designed specifically for geometry. It was generated autonomously by an unreleased, general-purpose reasoning model developed by OpenAI. The AI system not only found a counterexample to Erdős's conjecture but did so by building an unexpected bridge between two historically distinct fields of mathematics. Mathematicians who reviewed the work described the discovery as a milestone, comparing its creative leap to AlphaGo's famous 'Move 37' against Lee Sedol, signaling a new era in machine reasoning.[1][3][5][6]

To understand the magnitude of the breakthrough, one must understand the unit distance problem itself. For small numbers of points, the puzzle is trivial to solve by hand. If you have three points, arranging them into an equilateral triangle yields three pairs that are exactly one unit apart. Four points can be arranged into a rhombus made of two conjoined equilateral triangles, yielding five unit-distance pairs. But as the number of points, denoted mathematically as 'n', grows into the thousands or millions, calculating the maximum number of unit-distance pairs becomes extraordinarily complex.[3][4][7]

Erdős theorized that the most efficient way to maximize these pairs was to arrange the points in a regular, square-like grid. Based on this grid structure, he conjectured that the number of unit-distance pairs would grow only slightly faster than the number of points themselves—a nearly linear relationship. For 80 years, this upper bound was treated as received wisdom. Researchers believed that while they might refine the exact constant, the grid-based ceiling was fundamentally correct.[2][5][7]

The OpenAI reasoning model proved that the entire discipline had been looking in the wrong place. When researchers fed the conjecture into the model and instructed it to either prove or disprove the statement, the AI bypassed the traditional tools of discrete combinatorial geometry. Instead, it reached into a completely different mathematical domain: algebraic number theory. By utilizing infinite class field towers and a concept known as Golod-Shafarevich theory, the model constructed an infinite family of point arrangements that vastly outperformed Erdős's grid.[1][4][6]

The AI's construction demonstrated that the number of unit-distance pairs could actually grow at a significantly larger rate, mathematically defined as n^(1+δ), where δ is a fixed positive constant. This polynomial improvement definitively broke the conjecture. The grid was not the ceiling; it never was. The model had discovered a richer symmetry structure within algebraic number fields that allowed for far more points to sit exactly one unit apart than human intuition had ever anticipated.[1][5][6][7]

What makes this achievement historic is the nature of the reasoning involved. Previous AI milestones in mathematics often relied on raw computational brute force or pattern-matching against existing databases of known proofs. This problem, however, required generating novel ideas and maintaining logical coherence across a long horizon of dependencies. The AI had to hold multiple constraints in its memory, plan several steps ahead, and execute a flawless 125-page argument without drifting off-track.[1][2][3]

Furthermore, the cross-disciplinary nature of the proof stunned experts. Discrete geometry and algebraic number theory are highly specialized, separate cultures within mathematics. A human expert in one field might possess a passing familiarity with the other, but rarely at the depth required to synthesize a breakthrough of this magnitude. The AI's ability to seamlessly translate a geometric problem into an algebraic framework demonstrated a capacity for lateral, cross-domain thinking that had previously been considered the exclusive domain of human genius.[1][3][6]

The mathematical establishment quickly mobilized to verify the AI's claims. Thomas Bloom, a mathematician who maintains a database of Erdős's unsolved problems, and Melanie Matchett Wood of Harvard University were among the experts who reviewed the proof. They confirmed that the underlying mathematics was not only correct but aesthetically beautiful. The AI had succeeded by persevering down complex algebraic paths that human researchers had likely dismissed as too tedious or unpromising to explore manually.[1][2][6]

Human-AI collaboration immediately followed the initial discovery. While the AI's proof established that a positive constant δ existed, it did not explicitly calculate its optimal value. Princeton mathematician Will Sawin took the AI's foundational algebraic construction and optimized the Golod-Shafarevich step, sharpening the result to prove that δ could be at least 0.014. This rapid refinement highlighted what many researchers see as the future of the discipline: AI systems handling computationally intensive, cross-domain reasoning, while human mathematicians provide the judgment, optimization, and contextual understanding.[3][5][7]

Despite the celebration, the breakthrough has sparked a vigorous debate about transparency, trust, and the need for guardrails in AI-driven science. The unit distance problem is not entirely 'solved'—the AI merely disproved Erdős's upper bound, but the exact maximum rate of growth remains an open question for mathematicians to resolve. More importantly, the model's success was not absolute. OpenAI researchers revealed that when they ran the prompt through the model multiple times, it produced the correct, flawless proof in only 50 percent of the trials, highlighting the probabilistic nature of current AI reasoning.[2][4][7]

In the other 50 percent of attempts, the model either failed to solve the problem or generated confident-sounding but mathematically flawed reasoning. Skeptics and ethicists point out that OpenAI does not publicly share the model's failures or the full extent of its hallucination rates on complex formal logic. If an AI generates a 125-page proof that is subtly incorrect, verifying the errors could drain thousands of hours of human labor. Without knowing the probability of correctness upfront, mathematicians risk being buried in plausible but nonsensical AI-generated literature.[2][3][4]

There is also the persistent issue of data contamination. While OpenAI asserts that the model reasoned autonomously, the tools it used—algebraic number fields and Golod-Shafarevich theory—were part of its vast training data. The AI did not invent new mathematics from scratch; it found an unprecedented way to combine existing human knowledge. As AI systems are deployed against even harder, truly novel problems, their ability to generate entirely new mathematical frameworks remains untested.[3][4][6]

Nevertheless, the disproof of Erdős Conjecture #90 marks a permanent shift in the landscape of formal reasoning. It provides concrete evidence that general-purpose AI models are crossing the threshold from useful assistants to capable research partners. By breaking an 80-year-old ceiling, the AI has not replaced human mathematicians, but it has irrevocably expanded the toolkit available to them, proving that the answers to our oldest questions might lie in connections we simply couldn't see.[1][3][6]