An AI Model May Have Just Overturned Erdős’s Famous Unit Distance Conjecture

The unit distance problem explained

Imagine placing n points anywhere in the plane. Some pairs of those points may lie exactly one unit apart.

Mathematicians define:

• ν(P) — the number of pairs of points in a set P that are exactly distance 1 apart
• ν(n) — the maximum possible value of ν(P) among all point sets with n points

The question is simple to state but difficult to answer: how quickly can ν(n) grow as n increases?

Even though the problem sounds geometric, it sits at the crossroads of geometry, combinatorics, and number theory.

Erdős’s 1946 conjecture

Erdős proposed a construction based on arranging points roughly in a √n × √n grid. This arrangement produces about

n^(1 + Ω(1 / log log n))

pairs of points at distance 1. He conjectured that this was essentially optimal—meaning the number of unit distances should grow only slightly faster than linear in n.

In other words, Erdős believed the true growth rate should be close to

n^(1 + o(1)).

Meanwhile, mathematicians proved an upper bound showing that

ν(n) = O(n^(4/3)).

This bound, established by Spencer, Szemerédi, and Trotter in 1984, left a large gap between the best known lower and upper limits. Closing that gap became a famous challenge in discrete geometry.

The AI‑generated counterexample

The new work claims a dramatically different result.

According to the OpenAI proof, there exist infinitely many values of n for which

ν(n) ≥ n^(1+δ)

for some fixed positive constant δ.

That single change—from a slowly growing exponent to a fixed polynomial improvement—directly contradicts Erdős’s conjecture. If correct, it means that certain point configurations can generate far more unit‑distance pairs than mathematicians expected.

A surprising mathematical strategy

Earlier constructions focused mainly on geometric patterns, such as grids or lattice arrangements.

The new approach instead draws on deep algebraic number theory. At a high level, the proof builds special mathematical structures using:

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

These ingredients create high‑dimensional lattices containing many elements of norm 1. When translated into geometric configurations in the plane, those algebraic relationships correspond to large numbers of unit‑distance edges between points.

The shift from geometric intuition to number‑theoretic machinery is part of what makes the construction unusual—and powerful.

Why the result disproves the conjecture

Erdős’s prediction allowed only a tiny improvement over linear growth in the number of unit distances.

But the new construction produces a bound of

n^(1 + δ)

with δ fixed and positive.

Because the exponent does not shrink as n grows, the difference between the two predictions eventually becomes arbitrarily large. That makes the new construction a direct counterexample to the conjectured near‑linear bound.

Human mathematicians checked the argument

After the AI‑generated proof was produced, a group of mathematicians prepared a human‑verified presentation and commentary explaining the argument and its implications.

The authors include prominent researchers such as Noga Alon, Timothy Gowers, Thomas Bloom, Will Sawin, Melanie Matchett Wood, Daniel Litt, Arul Shankar, Jacob Tsimerman, and Victor Wang. Their document distills the core ideas of the construction and connects them to existing work in algebraic number theory.

Their involvement reflects how new mathematical results—especially surprising ones—are typically scrutinized and reformulated by experts before broader acceptance.

Why the discovery is attracting attention

The claim has sparked widespread interest because it appears to represent something rare: an AI system generating a genuinely new proof for a major open mathematical conjecture rather than rediscovering a known solution.

Previous experiments with AI‑generated mathematics sometimes produced correct answers that already existed somewhere in the literature. In this case, the result is presented as a new counterexample to a conjecture that stood since 1946.

If the mathematical community ultimately confirms the proof, the episode could mark a notable shift in how mathematical discovery happens—combining automated reasoning systems with traditional human verification.

What happens next

Results of this scale typically undergo months—or years—of careful examination.

Mathematicians will likely:

• check every step of the argument in detail
• search for simpler explanations or alternative constructions
• explore implications for related problems in combinatorics and geometry

Regardless of the final outcome, the work highlights a new possibility: AI systems helping researchers explore mathematical territory that was previously too vast to search systematically.