In mid-May 2026, OpenAI announced that one of its internal AI models had disproved the Erdős unit distance conjecture — a discrete geometry problem that had resisted human mathematicians for 80 years. Several leading mathematicians reviewed the result before publication, and their reactions were enthusiastic but measured.
Key takeaways
- The AI model disproved — not proved — the Erdős conjecture, constructing a point configuration with more unit distances than the conjecture allowed
- Fields Medal winner Tim Gowers called the result "a milestone in AI mathematics"
- The AI applied techniques from algebraic number theory and high-dimensional lattices — tools non-obvious in the context of this problem
- Mathematician Will Sawin showed the AI's result implies a lower bound of at least n^1.014 — higher than Erdős predicted
- The same week, Google announced its AI system solved nine open Erdős problems
- GPT-5.5 can disprove the conjecture with a small hint — a finding that is itself surprising
The unit distance problem
In 1946, Hungarian mathematician Paul Erdős posed a deceptively simple question: how many pairs of points placed in a 2D plane can be exactly distance 1 apart? More formally, he sought upper and lower bounds on this count as the number of points n grows large. Using a square grid as his construction, Erdős established that the optimal number of unit distances grows more slowly than n² but slightly faster than n.
For 80 years, mathematicians assumed Erdős was right about the direction. The best known lower bound was n^(1 + C/log log n); the upper bound stood at approximately n^1.333. Erdős conjectured that the true value was close to the lower bound — that his grid was nearly optimal.
The OpenAI AI model proved otherwise.
How the AI disproved the conjecture
Rather than proving the conjecture, the model constructed a counterexample: an arrangement of n points yielding more unit distances than the conjecture predicted.
The key insight was to build a lattice in high-dimensional space using algebraic integers, then project that structure down to two dimensions. This kind of lattice has richer mathematical structure than Erdős's square grid, allowing more unit-distance pairs to be packed into the same number of points.
Mathematician Will Sawin, one of the reviewers, subsequently showed that the AI's construction implies a lower bound of at least n^1.014 — clearly above the previous best. The result does not fully resolve the problem, however: the upper bound of n^1.333 remains. The gap between n^1.014 and n^1.333 still needs to be closed.
Why AI, and not a human mathematician
The obstacle was not a lack of ideas, but a strong prior about direction. Most mathematicians assumed Erdős was correct and looked for a proof of the conjecture, not a counterexample. Professor Jacob Tsimerman of the University of Toronto admitted he had briefly considered a similar approach to disprove it — but abandoned it as too laborious and unlikely to succeed. The expected cost of the search outweighed the anticipated payoff.
An AI system faces no such hesitation. It can methodically work through proof strategies that seem unpromising to a human. Notably, OpenAI's own data shows that even at maximum token budget, the model solved the problem only half of the time. The result was reproducible, but not trivial.
The second factor was breadth of knowledge. The algebraic techniques the model applied come from a field far removed from discrete geometry. AI systems trained on vast bodies of mathematical literature can draw connections across distant subfields — a capability that human experts, focused by specialization, are structurally less likely to exploit.
Context: what Google was doing at the same time
Two days after OpenAI's announcement, on May 22, Google reported that its AI system had solved nine open Erdős problems, including two that had been unsolved for over 50 years. This is a separate achievement using a different technical approach, but it confirms that multiple laboratories are simultaneously probing Erdős's catalog as a benchmark for autonomous AI reasoning.
Earlier, Google DeepMind's AlphaEvolve had drawn attention as a system that uses language models as an optimization engine. When a mathematical problem can be cast as code to optimize, AlphaEvolve can find solutions better than existing human results — as confirmed by four research teams analyzing its performance on 67 optimization problems in November 2025. OpenAI went a step further: their model did not optimize an existing formulation, but autonomously constructed a full mathematical proof.
Why this matters
The results from OpenAI and Google settle the question of whether language models can contribute to the production of new mathematical knowledge. The answer is yes — and not merely as assistants, but as autonomous resolvers of open problems.
At the same time, precision matters here. The AI did not invent a fundamentally new mathematical technique. It combined existing tools in a non-obvious way. The proof required human verification and extension. And the result closes one question while leaving a gap between n^1.014 and n^1.333 that future work must address.
More than a single breakthrough, OpenAI's result signals a shift in dynamics: AI can now independently identify directions that mathematicians later develop. Just a year ago, language models were first clearing high school mathematics competitions. Today they contribute to research mathematics. That is a faster progression than most mathematicians anticipated.
What's next?
- The gap between the AI's lower bound of n^1.014 and the upper bound of n^1.333 remains open — closing it is the natural next target for both human and AI researchers
GPT-5.5, a publicly available OpenAI model, can disprove the conjecture with a small hint — discovered by PhD student Xiao Ma shortly after the announcement; OpenAI has indicated it will investigate how many other known problems are within reach of current models
- The Erdős problems list, maintained by mathematician Thomas Bloom at erdosproblems.com, is becoming a de facto benchmark for autonomous mathematical AI — other laboratories have signaled their own campaigns against selected entries on the list
Sources
- Ars Technica — An OpenAI model solved a famous math problem that stumped humans for 80 years
- OpenAI — Model disproves discrete geometry conjecture
- arXiv / Will Sawin — Extensions of the unit distance result
- arXiv / OpenAI — Unit distance problem — mathematicians' remarks
- Google Research — AI solves nine open Erdős problems
