# OpenAI model disproves a longstanding unit-distance conjecture > An internal general-purpose reasoning model found an infinite family of point configurations; external mathematicians checked the resulting proof. Canonical URL: https://www.devobs.io/news/news-openai-unit-distance-conjecture-disproof/ By: Owen Park Published: 2026-09-06T11:58:54.623Z Updated: 2026-09-06T11:58:54.623Z Event date: 2026-05-20 Section: AI OpenAI announced [a disproof of a central conjecture in the planar unit-distance problem](https://openai.com/index/model-disproves-discrete-geometry-conjecture/) on May 20. An internal general-purpose reasoning model produced an infinite family of point configurations with a fixed polynomial improvement over the long-believed near-linear construction bound. OpenAI says external mathematicians checked the proof and published companion remarks about its significance. ## The result changes the lower-bound picture The problem asks for the maximum number of unit-distance pairs among a set of points in the plane. For decades, a rescaled square-grid construction supported the belief that the best lower bound grew as n raised to 1 plus a term tending to zero. The new construction yields at least n to the 1 plus delta unit distances for infinitely many n, for some fixed positive delta, disproving that belief. The argument unexpectedly imports techniques from algebraic number theory into discrete geometry. The discovery process is also notable. OpenAI says the model was neither specialized for mathematics nor specifically scaffolded for this problem; it was tested on a collection of Erdős problems and pursued a counterexample. The published proof and independent companion paper are more important evidence than a model transcript because mathematical validity rests on an inspectable argument. ## Verification remains the durable interface This result is stronger than a plausible conjecture or numerical search, but it does not imply autonomous reliability across open mathematics. Researchers should separate problem selection, candidate generation, proof checking, formalization, and interpretation. Each stage has different failure modes. For AI-assisted research, preserve the exact problem statement, model and harness version, candidate proof, external reviews, and later corrections. The useful workflow is asymmetric: models can search broadly and connect distant fields, while specialists decide whether the argument is correct, novel, and meaningful. ## Source references -