OpenAI says model disproved Erdős unit‑distance conjecture

On May 20, 2026 OpenAI posted that one of its internal general‑purpose reasoning models produced an original mathematical construction that disproves a central conjecture in discrete geometry known as the planar unit distance problem. The company published the proof and companion remarks online.

The planar unit distance question asks, for n points in the plane, how many pairs can be exactly one unit apart. Paul Erdős posed the problem in 1946 and for decades many researchers believed square‑grid style constructions were essentially optimal.

OpenAI says the model produced an infinite family of point sets achieving at least n^{1+δ} unit‑distance pairs for some fixed δ>0, overturning the widely held conjecture that the maximum grows only like n^{1+o(1)}. The firm says external mathematicians checked the proof and produced a companion human‑readable note.

Technically the AI argument moves the problem into algebraic number theory: OpenAI’s writeup points to constructions built from number fields, infinite class‑field towers and Golod–Shafarevich methods to manufacture many unit‑length differences. That connection surprised many researchers because the problem looks purely Euclidean at first glance.

OpenAI’s initial AI‑generated argument did not pin down a concrete value for δ. Within hours and days of the announcement Princeton mathematician Will Sawin posted a short paper that refines the construction and gives an explicit exponent δ≈0.014, showing sets with roughly n^{1.014} unit pairs for infinitely many n.

The company made multiple artifacts public: a full proof PDF, a companion paper by nine human mathematicians that digests and vouches for the argument, and an abridged version of the model’s chain of thought used to find the construction. Those materials are intended to let specialists verify and reproduce the work.

OpenAI’s announcement rekindled memories of a public misstep last year, when a high‑level post implied GPT‑5 had solved several Erdős problems but the result mostly reflected rediscovery of existing literature. Reporters note that this time leading figures including Tim Gowers and Noga Alon are co‑signers or endorsers of the human companion note, which lends credibility but does not end scrutiny.

Verification and reproducibility are now central questions. The companion note is a human‑verified exposition of the AI construction, but mathematicians emphasize formal checking, independent re‑derivation, and publication in peer‑reviewed journals remain the normal bar for acceptance. The arXiv uploads open the work to that community inspection.

A separate strand of debate concerns the AI itself. OpenAI frames the event as the first time a general‑purpose reasoning model autonomously solved a field‑centred open problem rather than a system specifically tuned for math benchmarks. That claim is consequential for how researchers treat models as creative partners. Skeptics stress that human verification, interpretability of the reasoning chain, and reproducible methods matter more than an initial headline.

Prominent mathematicians who reviewed the material have given measured praise. Tim Gowers, in the companion paper, called the result “a milestone in AI mathematics” and said he would have recommended the write‑up for a top journal if a human had authored it. Others pointed out the novelty lies both in the construction and in the unexpected number‑theoretic bridge it reveals.

Practically, the community will now test limits and consequences. Researchers will try to simplify or generalize the construction, search for other geometry problems where number‑theory helps, and attempt fully formalized verifications inside proof assistants. The Sawin refinement gives a concrete numeric target that makes independent checks easier.

If the construction and proofs survive extended scrutiny, the episode will mark a clear example of AI moving beyond assistance toward originating competitive research ideas. That shift raises questions about credit, reproducibility standards, how to archive chains of thought, and what sort of governance or disclosure is needed when models claim new facts. For now mathematicians and AI researchers alike are poring over code, constructions and arXiv notes.