AI Disproves Conjectures: How Counterexamples Work

Quick answer: AI can disprove a universal mathematical conjecture by finding a counterexample: an object that satisfies the conjecture’s assumptions but breaks its conclusion. AI guides the search, generates candidate objects or programs, and develops constructions; rigorous checking then establishes the result. Examples include Adam Zsolt Wagner’s 2021 graph counterexamples, PatternBoost’s 2024 hypercube result, and the 2026 disproof of Erdős’s unit distance conjecture, which needed an infinite family of point configurations, not just one example.

Updated November 22, 2026.

Proving that a statement is true for every case can take centuries. Proving it false takes just one example. That asymmetry makes counterexample hunting one of the places where AI has been most useful in research mathematics, and where its results are easiest to check.

Why One Counterexample Is Enough

A conjecture such as “every \(X\) has property \(P\)” fails if you can exhibit a single \(X\) without \(P\). A simple example: the claim “\(n^2 + n + 41\) is prime for every whole number \(n\)” holds for \(n = 0\) to \(39\), but at \(n = 40\) it gives \(40^2 + 40 + 41 = 1{,}681 = 41^2\). One value breaks it.

A counterexample is usually easier to verify than to find: anyone can check it without trusting how it was found, though verification can still need serious computation or proof. For AI, that is ideal. The hard part is the search, because the space of possible objects can be astronomically large. Note the difference between a disproof and an improved bound: AlphaEvolve’s 593 spheres raised a record, but disproved nothing.

Computer-Assisted Disproofs Before AI

ConjectureDisprovedCounterexample
Euler’s sum of powers conjecture1966, Lander and Parkin\(27^5 + 84^5 + 110^5 + 133^5 = 144^5\)
Pólya conjecture1958 (Haselgrove); smallest counterexample found 1980Fails first at \(n = 906{,}150{,}257\)
Mertens conjecture1985, Odlyzko and te RieleProved false without an explicit example

These show a warning that applies to AI too: a pattern that holds for millions of cases can still fail. Checking small cases is evidence, not proof.

How AI Searches for Counterexamples

  1. Turn the conjecture into a score. For “every graph satisfies \(A \ge B\),” score a graph by \(B – A\). A positive score is a counterexample.
  2. Generate candidates. A neural network or language model proposes objects, or programs that build objects.
  3. Keep the best. The top-scoring candidates are used to train or prompt the next round.
  4. Verify exactly. Any claimed counterexample is checked with exact arithmetic or a proof.

AI Counterexamples: A Timeline

YearResultMethod
2021Adam Zsolt Wagner refutes several conjectures, including a question of Brualdi and Cao and problems on graph eigenvaluesReinforcement learning (deep cross-entropy method)
2023FunSearch finds larger cap sets than previously known in some dimensionsLanguage model plus program search
2024PatternBoost (François Charton, Jordan Ellenberg, Wagner, Geordie Williamson) finds a counterexample to a 30-year-old conjecture about hypercubes: a spanning subgraph of the 6-cube with diameter 6 and only 81 edges, beating the conjectured minimum of 82Local search plus a transformer
2025AlphaEvolve finds a 593-sphere kissing arrangement in 11 dimensions (an improved bound, not a disproof)Evolutionary program search
May 2026Erdős’s 1946 unit distance conjecture disproved with an infinite family of point sets; proof checked by outside mathematiciansOpenAI model
July 2026Jacobian conjecture fails in three (and more) variables; the two-variable case remains openCounterexample credited to Levent Alpöge working with Claude; independently verified in Isabelle

Example: Wagner’s Graph Counterexamples

In his 2021 paper “Constructions in combinatorics via neural networks,” Wagner had a neural network build graphs edge by edge. Each finished graph was scored by how badly it violated a conjecture, and the network learned from the best graphs. It found counterexamples to several published conjectures about graph eigenvalues, often strange-looking graphs that people had not thought to try. The method is simple enough that many mathematicians reused it.

Example: The 2026 Breakthroughs

The unit distance problem asks how many pairs of points among \(n\) points in the plane can be exactly distance 1 apart. Erdős conjectured in 1946 that the maximum grows like \(n^{1+o(1)}\); in May 2026 an OpenAI model constructed an infinite family with at least \(n^{1+\delta}\) unit distances for a fixed \(\delta > 0\), so no single example would have sufficed. It does not determine the exact maximum. In July 2026, a three-variable polynomial map with constant nonzero Jacobian determinant that is not one-to-one disproved the Jacobian conjecture in dimension three and higher. In both cases, the counterexample could be checked independently. See Erdős problems solved by AI.

Limits of Counterexample Search

  • It only disproves. Failing to find a counterexample proves nothing.
  • Some counterexamples are huge. The Pólya conjecture first fails near 906 million; others may be far beyond any search.
  • The score must be honest. Rounding errors can make a non-example look like one, so final checks use exact arithmetic.
  • Understanding still matters. Mathematicians want to know why the counterexample works, which often leads to new theorems.

Common Misconceptions

“Checking a billion cases proves a conjecture.” No. Only a proof covers every case.

“AI counterexamples are hard to trust.” Usually the opposite: a concrete counterexample can be checked by anyone.

“AI only finds small examples.” Search methods can build very large objects, though some counterexamples may be out of reach.

Watch: AI and the Hunt for Counterexamples

If you learn better by watching, these videos are worth your time.

Geordie Williamson: Neural Networks for Mathematical Discovery (Simons Foundation). A PatternBoost coauthor on how neural networks help find mathematical objects (October 2025).
Has an AI discovered new maths? (Stand-up Maths). How to judge claims that AI has found something new.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. How many counterexamples does it take to disprove a conjecture?
Found a bug? Let us know!
2. What does an AI counterexample search need?
Found a bug? Let us know!
3. Which conjecture did AI disprove in May 2026?
Found a bug? Let us know!
4. Does failing to find a counterexample prove a conjecture?
Found a bug? Let us know!
5. AlphaEvolve found 593 touching spheres in 11 dimensions. What kind of result is that?
Found a bug? Let us know!

Frequently Asked Questions

How does AI disprove math conjectures?

It turns the conjecture into a score that measures how close a candidate object comes to breaking it, generates many candidates with neural networks or language models, learns from the best ones, and has any claimed counterexample checked exactly.

What conjectures has AI disproved?

Examples include several graph-theory conjectures (Wagner, 2021), a 30-year-old conjecture via PatternBoost (2024), Erdős’s unit distance conjecture (May 2026) and the Jacobian conjecture in three variables (July 2026).

Why are counterexamples easy to trust?

A counterexample is a concrete object that anyone can check, often by computer, without trusting how it was found.

Can checking many cases prove a conjecture?

No. A conjecture can hold for millions of cases and still fail; the Pólya conjecture first fails at 906,150,257. Only a proof covers every case.

What was the first computer counterexample?

A famous early one is Lander and Parkin’s 1966 computer search disproving Euler’s sum of powers conjecture: 27⁵ + 84⁵ + 110⁵ + 133⁵ = 144⁵.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AI Disproves Conjectures: How Counterexamples Work - Effortless Math"?

No one replied yet.

Leave a Reply