The AI‑Generated Counterexample to Erdős’s Unit Distance Conjecture

Erdős’s original conjecture

Erdős constructed point sets arranged roughly in a √n × √n grid, which produce about

n^(1 + Ω(1 / log log n))

unit‑distance pairs. He conjectured this was essentially optimal—meaning ν(n) should be close to linear, with no fixed power improvement beyond n.

Meanwhile, a major upper bound proved by Spencer, Szemerédi, and Trotter (1984) showed that

ν(n) = O(n^(4/3)).

This left a wide gap between the best known lower and upper bounds, and closing it became a central open problem in discrete geometry.

The AI‑generated counterexample

The OpenAI work constructs families of planar point sets that achieve

ν(n) ≥ n^(1+δ)

for some fixed δ > 0 and infinitely many values of n.

Because Erdős’s conjecture predicted that ν(n) could grow only slightly faster than linear (n^(1+o(1))), the existence of any fixed exponent improvement directly contradicts the conjecture.

In other words, the new construction shows that certain configurations of points can create polynomially more unit distances than previously believed possible.

The key mathematical idea: number theory instead of grids

Earlier constructions relied mostly on geometric or lattice‑based patterns such as grids. The new approach takes a very different route: algebraic number theory.

At a high level, the proof uses several advanced ingredients:

• Totally real number fields arranged in infinite class field towers with special splitting properties
• Golod–Shafarevich–type constructions, which guarantee infinite towers of number fields with controlled arithmetic structure
• CM fields obtained by adjoining the imaginary unit i

These structures produce high‑dimensional lattices with many elements of norm 1. When the construction is mapped into the Euclidean plane, those norm‑one relations correspond to large numbers of unit‑distance edges between points.

The key advantage is that number‑theoretic constructions generate far richer families of distance relations than the classical grid approach.

Why this breaks the long‑standing conjecture

The difference between the two growth rates is subtle but decisive.

Erdős’s conjecture allowed only

n^(1 + o(1))

unit distances.

The new construction produces

n^(1 + δ)

for a fixed positive δ. Because the exponent improvement does not shrink with n, the gap eventually becomes arbitrarily large—making the conjecture false.

Human verification by mathematicians

After the result was produced, mathematicians examined the argument and published a short, human‑verified presentation of the proof along with commentary.

The verification document involves researchers including Noga Alon, Timothy Gowers, Thomas Bloom, Will Sawin, Melanie Matchett Wood, and others, who summarized the argument and discussed its ideas and implications.

Their analysis explains how the construction connects several strands of number theory—including work related to Golod–Shafarevich towers and related algebraic techniques—to generate the geometric configurations underlying the counterexample.

Why the result is being called a milestone for AI

The breakthrough has attracted attention because it appears to represent an AI system generating a new proof for a prominent open mathematical conjecture, rather than rediscovering a known solution.

That distinction matters. Earlier experiments with AI‑generated mathematics sometimes produced correct solutions that later turned out to already exist in the literature. In this case, the reported result is presented as a genuinely new counterexample to a conjecture that had remained open since 1946.

If the mathematical community ultimately accepts the proof as correct and complete, it would mark a notable moment in computational mathematics: an AI system contributing a novel insight to a major unresolved problem in discrete geometry.

What happens next

Even with verification summaries available, results of this scale typically undergo extensive scrutiny by the mathematical community. Researchers will examine the details, attempt to simplify the arguments, and explore the implications for related geometric and combinatorial problems.

Regardless of the final verdict, the work illustrates a shift in how mathematical discovery can happen—combining automated reasoning systems with human verification to explore parts of the mathematical landscape that were previously difficult to search systematically.