OpenAI Model Disproves 80-Year-Old Erdős Geometry Conjecture in Verified Breakthrough

By Hector Herrera | May 21, 2026 | Science

An OpenAI reasoning model has disproved a central conjecture in discrete geometry first posed by mathematician Paul Erdős in 1946 — and the result has been independently verified by external mathematicians. This is the first time an AI has autonomously resolved a prominent open problem at the frontier of a mathematical subfield, not by assisting a human researcher but by finding the proof on its own.

The result matters because it moves AI from a tool that helps mathematicians to one that can advance mathematics.

Background

Paul Erdős was one of the most prolific mathematicians of the 20th century, famous for posing deceptively simple-sounding problems that resisted solution for decades. The conjecture in question concerned discrete geometry — the study of geometric objects defined by discrete (integer or combinatorial) structures, as opposed to continuous curves. For nearly 80 years, the square-grid arrangement of points had been considered the optimal construction for the problem, and the field assumed the conjecture was likely true.

AI systems have previously assisted with mathematical reasoning — most notably DeepMind's AlphaProof, which tackled International Mathematical Olympiad problems. But those were structured competition problems with known solution types. An open conjecture at the frontier of a subfield is a different challenge.

What Happened

According to OpenAI's announcement, a general-purpose reasoning model — not a specialized math system — identified an infinite family of counterexamples that outperform the square-grid constructions believed optimal since 1946. The model applied techniques from algebraic number theory (the study of number fields and their arithmetic properties) to construct these examples, a methodological crossover that human researchers had not attempted in this context.

Key details:

• Problem origin: Erdős conjecture on point arrangements in discrete geometry, posed 1946
• Previous best construction: Square-grid arrangements, held as optimal for ~80 years
• AI's contribution: Identified an infinite family of arrangements that improve on the square grid
• Methods used: Algebraic number theory techniques applied to a discrete geometry problem
• Verification: External mathematicians confirmed the result independently

OpenAI has not specified which model produced the result or the exact computational resources involved.

Why This Is Different From Prior AI Math Results

Most AI math achievements to date — including performance on the Putnam exam and IMO problems — involve solving problems that have known answers being tested against a rubric. Disproving an open conjecture is categorically different: the AI had to identify that the accepted belief was wrong, construct a counterexample, and do so in a domain where no formal benchmark existed.

The use of algebraic number theory to attack a discrete geometry problem also suggests the model was not pattern-matching against known proof strategies for this specific type of problem. It appears to have synthesized techniques across subfields — something mathematicians spend careers learning to do.

Impact

For mathematics: Open problems are the currency of the field. Resolving even one — especially one carrying the Erdős name — is significant. The question now is whether AI can do this systematically or whether this was an exceptional case.

For AI development: This result raises the bar for what counts as a meaningful AI capability benchmark. If general-purpose reasoning models can autonomously advance a scientific field, the implications extend well beyond mathematics — to drug discovery, materials science, and any domain defined by unresolved theoretical questions.

For research institutions: Universities and research labs will need to reckon with a new question: when an AI resolves a long-standing problem, who gets credit, and how does that reshape the incentive structure of academic research?

What to Watch

OpenAI has not yet published a peer-reviewed paper on the result. The next step is formal mathematical publication with full proof verification — which will determine whether this is a reproducible method or a one-time result. Watch for whether other labs attempt to replicate the approach on other Erdős problems, of which hundreds remain open.

Sources: OpenAI announcement