OpenAI Model Disproves 80-Year-Old Erdős Unit Distance Conjecture

An OpenAI model has disproved the planar unit distance conjecture — a famous unsolved problem in mathematics that Paul Erdős first posed nearly 80 years ago. OpenAI announced the result on 20 May 2026, and the finding has since been independently checked by human mathematicians.

The conjecture concerns a deceptively simple question: given *n* points scattered on a flat plane, what is the maximum number of pairs that can sit exactly one unit apart? For decades, the assumption held that this number couldn’t grow much faster than a certain known bound. The OpenAI model found a sequence of point sets that breaks that assumption entirely.

What the Model Actually Found

The OpenAI paper states there exists a value ε greater than zero, and a sequence of point sets in the plane where the number of unit-distance pairs grows at least as fast as |P|¹⁺ᵋ — where |P| is the number of points. That’s faster than anyone had previously proved possible.

To put that in plain terms: the model found arrangements of points that produce more same-distance pairs than the old theoretical ceiling allowed. That ceiling — the best known upper bound, cited in the follow-up arXiv note as O(n⁴/³) — had stood unchallenged for years. Erdős himself had shown a lower bound construction of n¹⁺Ω(1/log log n), and for a long time grid-like arrangements of points were considered the best anyone could do.

The tools involved are not simple. According to the OpenAI paper and associated reporting, the proof draws on algebraic number theory — specifically infinite class field towers and Golod-Shafarevich theory. These are not household names even in mathematics departments, which gives some sense of how deep the result goes.

Princeton Mathematician Refines the Result

Will Sawin, a mathematician at Princeton University, then refined the finding. His contribution, linked in the OpenAI-associated reporting and the arXiv remarks, showed the improvement could be expressed using a fixed exponent rather than the more open-ended formulation in the original result.

That kind of follow-on refinement is standard in mathematics. A first proof establishes that something is true; later work sharpens exactly how true it is.

The arXiv note describes the result as “human-verified” — meaning mathematicians reviewed the model’s output and confirmed the counterexample holds. That verification matters. It separates this from a model producing plausible-sounding mathematics that later falls apart under scrutiny.

A Genuine Breakthrough — With Questions

OpenAI is presenting the work as evidence that AI can assist in genuine frontier mathematical discovery, not just crunch numbers or search databases.

That claim carries weight here. The Erdős unit distance problem isn’t an obscure puzzle. It sits at the heart of combinatorial and discrete geometry, and it attracted serious attention from serious mathematicians — including Larry Guth and Nets Katz, who made progress on related Erdős problems in prior years. The fact that it resisted solution for nearly eight decades says something.

But some researchers are likely to ask how much of the discovery was model-led and how much came from human guidance, refinement, and the selection of mathematical tools. The line between “AI found this” and “AI helped a mathematician find this” isn’t always clear, and the answer matters for how the field interprets what happened.

OpenAI hasn’t claimed the model worked entirely alone. The framing has been one of AI-assisted mathematical research — a collaboration rather than a replacement.

Why Erdős Still Matters

Paul Erdős, the Hungarian mathematician who posed this problem in 1946, was one of the most prolific mathematicians of the 20th century. He published around 1,500 papers and collaborated with hundreds of researchers, often offering small cash prizes for solutions to problems he considered hard but tractable.

The unit distance problem was one of his. He didn’t live to see it resolved — he died in 1996.

His name still attaches to dozens of open problems across graph theory, number theory, and combinatorics. Disproving one of them, in the way OpenAI describes, would be a notable moment in the history of the subject regardless of who — or what — did it.

What This Means for Kent Residents

There’s no direct day-to-day impact on people in Kent from this result — it’s a development in pure mathematics, not a product launch or policy change. However, schools and universities across the county may well point to this as a real-world example of AI being used in advanced research, which is increasingly relevant for students considering careers in mathematics, computing, or science. More broadly, it adds to a growing picture of AI tools moving beyond writing and image generation into areas that were once thought to require exclusively human reasoning.