OpenAI announced on May 20, 2026, that an internal general-purpose reasoning model had produced a counterexample to Paul Erdős's 1946 unit distance conjecture — a result in discrete geometry that had resisted eight decades of human effort. Unlike the company's discredited October 2025 announcement, which collapsed within days under mathematical scrutiny, this one arrived with a companion verification paper co-authored by nine external mathematicians, including Thomas Bloom, the very researcher who publicly exposed the earlier false claim. Fields Medalist Tim Gowers, writing in the companion paper, said he would recommend the result for acceptance in the Annals of Mathematics — one of the field's most selective journals — without hesitation.
October 2025: When OpenAI's Math Talk Outran Its Mathematics
In October 2025, then-OpenAI vice president Kevin Weil posted on X that GPT-5 had found solutions to ten previously unsolved Erdős problems and made progress on eleven others. The post was deleted within days. Bloom, who maintains the Erdős Problems database at erdosproblems.com, explained that the model had not produced original proofs — it had retrieved existing solutions already buried in the mathematical literature. Bloom called the announcement "a dramatic misrepresentation." Google DeepMind CEO Demis Hassabis called it "embarrassing." Weil left OpenAI in April 2026. When Bloom appears now as a co-author of the verification paper for the new result, the credibility context changes materially.
What the Conjecture Claims, and What the Proof Shows
The unit distance problem asks a deceptively simple question: given n points scattered in the plane, how many pairs can sit exactly one unit apart? Erdős conjectured in 1946 that the count can grow no faster than n^(1+o(1)) — a rate only barely faster than linear. For nearly eight decades, no construction beat that near-linear ceiling by a fixed polynomial exponent, and the best-known upper bound of O(n^(4/3)), established by Spencer, Szemerédi, and Trotter, has stood since 1984.
OpenAI's model proved the ceiling is wrong. Its construction demonstrates point arrangements that yield at least n^(1+δ) unit-distance pairs for some fixed positive exponent δ — a polynomial improvement, not merely a sub-polynomial one. Princeton mathematician Will Sawin subsequently refined the result, pinning the explicit exponent at δ = 0.014. That figure is small, but the qualitative change — from sub-polynomial to polynomial growth — is exactly what makes Erdős's conjecture wrong rather than merely imprecise.
How the Proof Reaches from Geometry into Number Theory
The route the model took is, mathematically, the strangest part of the result. Erdős's original lower-bound construction can be understood through the Gaussian integers — complex numbers with integer parts. The new proof replaces these with more elaborate algebraic number fields possessing richer symmetries, then uses infinite class field towers and Golod-Shafarevich theory to prove the required number fields actually exist. These are tools from deep algebraic number theory — far outside the standard discrete-geometry toolkit — that mathematicians had not previously connected to the unit distance problem.
Sébastien Bubeck, the OpenAI mathematician leading the company's mathematical research, described the outcome: "The model did not invent something fundamentally new that nobody saw coming. It just executed like an amazing mathematician." Jacob Tsimerman, a mathematician at the University of Toronto who was consulted to verify the proof, identified why an AI might succeed where humans had stalled: "AIs have an edge: It's not just that they can try all known methods. They can play for longer and in more treacherous waters than mathematicians without getting overwhelmed."
Who Ran the Model and What Happened to the Output
According to attribution in Will Sawin's follow-up paper, the proof was produced by OpenAI researcher Lijie Chen using an internal model, with Mark Sellke and Mehtaab Sawhney — both OpenAI mathematicians — verifying its correctness. The companion arXiv paper confirms that the underlying proof file was generated in one shot by the model, with exposition subsequently refined through human interaction with Codex.
One transparency caveat that mathematicians have flagged: no external reviewer has seen the model's original raw output. The verification and companion papers are based on an edited version of the AI's chain of thought. Thomas Bloom, in his section of the companion paper, wrote that "the human still plays a vital role in discussing, digesting, and improving this proof, and exploring its consequences." Melanie Matchett Wood of Harvard added a pointed observation: had the assembled human experts combined their efforts for the same time it took them to parse the AI's answer, she concluded, "the mathematicians would have found a counterexample." The tools existed; the approach, in hindsight, was not out of reach. The AI found it first.
What Nine Mathematicians Verified — and What They Said
The companion arXiv paper, posted May 20, 2026, lists Noga Alon, Thomas Bloom, W.T. Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood as authors. They describe their document as "a short, digested, human-verified version of the recent OpenAI-generated counterexample."
Noga Alon, a combinatorialist at Princeton who heard Erdős raise the unit distance problem personally in lectures, called the AI's solution "an outstanding achievement, settling a long-standing open problem," and noted that the construction applies algebraic number theory tools "in an elegant and clever way." Daniel Litt, at the University of Toronto, offered the clearest summary of where this result sits in the history of AI mathematical work: "This is the unique interesting result produced autonomously by AI so far." Gowers stated that "no previous AI-generated proof has come close" to meeting the standard for publication in a top mathematics journal.
What This Result Does and Doesn't Establish
Reasonable skepticism about AI mathematical claims has not expired. The model has not been publicly released, so independent reproduction of its approach on related problems is not yet possible. The published paper is the human-cleaned version; how much the polished argument diverges from the raw AI output will keep mathematicians debating for months. And Sawin's δ = 0.014 is a lower bound on the improvement exponent — an upper bound matching the new lower bound remains an open problem.
What the result does establish, by the assessment of the mathematicians who checked it, is the strongest documented case yet of an AI system autonomously settling a prominent open problem in pure mathematics. The framework for distinguishing that kind of claim from promotional noise now has a concrete standard: a named Fields Medalist willing to put the result in the Annals of Mathematics, co-signed by the researcher who caught the last false one.
In October 2025, OpenAI's mathematics talk ran ahead of its mathematics. This time the mathematics ran first.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
