Full Breakdown
OpenAI’s AI Model Disproves the Erdos Unit Distance Conjecture
6/3/2026, 2:34:36 AM
AI Disproves the Erdos Unit Distance Conjecture
In mid-May 2026 OpenAI announced its AI model generated ideas and a proof that disproved the Erdos unit distance conjecture, an 80-year-old problem. A team of nine mathematicians, including Tim Gowers, verified and published the result. The AI connected the geometry question to algebraic number theory via the Golod-Shafarevich criterion, producing point configurations whose unit-distance count grows as n^1.014.
Context: The Conjecture and Recent AI Advances
The Erdos unit distance conjecture, posed by Paul Erdos in 1946, asked how many pairs of points on a plane can be exactly one unit apart, with Erdos predicting growth. No proof or counterexample emerged for eight decades. AI’s role accelerated: three years ago large language models struggled with arithmetic; last year they won high-school contests; by early 2026 they produced contributions.
Key Actors
- OpenAI – Built the general-purpose language model that generated the proof ideas.
- Mathematicians – Nine researchers, including Tim Gowers and Daniel Litt, verified the proof.
- Google DeepMind – Launched AlphaProof Nexus a day later, solving nine Erdos problems with Gemini 3.1 Pro and the Lean assistant.
- Terence Tao – Warned of a shift toward “proof abundance.”
- Thomas Bloom, Yann LeCun, Demis Hassabis – Critiqued AI claims.
Data & Statistics
- Conjecture unsolved for 80 years.
- AI-generated point sets grow as n^1.014 in unit-distance pairs.
- AlphaProof Nexus solved 9 of 353 open Erdos problems (~2.5 %) and proved 44 OEIS conjectures.
- Each DeepMind solution cost a few hundred dollars in compute.
Official Statements & Responses
OpenAI said the AI supplied the core ideas but stressed human verification was essential. DeepMind’s CEO Demis Hassabis said the system is not yet AGI, framing it as a workflow tool rather than a step toward artificial general intelligence. Terence Tao warned the community must build infrastructure for the surge of AI-generated proofs.
Criticism & Opposition
OpenAI’s October 2025 claim that GPT-5 solved ten Erdos problems was called “a dramatic misrepresentation” by Thomas Bloom and mocked by Yann LeCun; Hassabis later called it “embarrassing.” Critics note the AI’s proof reused existing techniques and that counterexamples are less demanding than constructive proofs.
Conflicting Reports & Gaps
Some sources describe the AI’s work as a “creative breakthrough,” while others argue it merely recombined known ideas without pioneering new methods. Whether AI can autonomously produce deep, structurally novel proofs remains unverified.
Verbatim Quotes
- “there is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics.” — Tim Gowers, Fields Medalist
- “this is the first example of a result produced autonomously by an AI that I find exciting in itself, as opposed to as a leading indicator.” — Daniel Litt, University of Toronto
- “com database, called it “a dramatic misrepresentation.” — Thomas Bloom, mathematician
- “embarrassing.” — Demis Hassabis, CEO, Google DeepMind
What’s Next
OpenAI plans to embed the model in broader research tools, while DeepMind aims to expand formal verification pipelines. Both expect future AI to handle exploratory calculations, leaving humans to choose problems and interpret significance.
