OpenAI says internal model produced a proof refuting the Erdős unit‑distance conjecture

In mid‑May OpenAI announced that an internal AI model produced a proof that refutes the Erdős unit‑distance conjecture, a longstanding question in discrete geometry posed by Paul Erdős in 1946. The company provided early access to the result to several external mathematicians for review, and Fields Medalist Tim Gowers described the outcome as “a milestone in AI mathematics.” If verified, the finding would mark a notable moment for AI assistance in producing research‑level mathematics and could shift expectations about what models can contribute to open problems.

Commentary published after the disclosure says the model constructed a full proof by assembling existing ideas drawn from multiple mathematical subfields rather than inventing fundamentally new techniques. OpenAI’s output was then cleaned up and extended by human mathematicians, and the company shared the revised work with outside experts for scrutiny. University of Toronto professor Daniel Litt called the case “the first example of a result produced autonomously by an AI that I find exciting in itself.

The Erdős unit‑distance problem asks, for n points placed in the plane, how many distinct pairs of points can be exactly one unit apart; Erdős introduced the question in 1946. For small values of n researchers have enumerated optimal arrangements — Boris Alexeev, Dustin G. Mixon, and Hans Parshall have provided optimal solutions through 21 points — but combinatorial complexity grows quickly as n increases, which left the general conjecture open until this claimed disproof.

Observers place the episode in the context of rapid improvements in large language model mathematics. Three years ago LLMs typically struggled with basic arithmetic; by last year they were performing strongly on high‑school‑level math competitions. Participants at the Joint Mathematics Meetings in January reported that AI systems were beginning to contribute to research, but usually in constrained settings and often requiring significant human interpretation to turn an AI output into a publishable theorem.

Technically, the episode highlights two model strengths relevant to research: broad recall of prior work across subfields and the willingness to grind through long, tedious proof strategies that humans might avoid. Humans remain essential, however, for deeper conceptual analysis, formal cleanup, and extending machine outputs into polished arguments. That hybrid workflow — model proposes, humans validate and refine — is evident in how the OpenAI output was handled after the initial disclosure.

For builders, verifiers, and platform teams the case underscores practical needs: reproducible proof artifacts, tooling to translate heuristic model outputs into rigorously checked proofs, and robust external review pipelines. It also poses a forward‑looking question about roles in the field: given the current pace of improvement, the role of human mathematicians could change substantially over the next decade, even as human oversight remains critical today.