Unsolved Math Problems Solved by AI

As of October 7, 2026, the evidence is strong that AI has materially advanced mathematics and has in some cases produced genuine solutions or disproofs of previously open problems—but the status is uneven, and the most dramatic claims are not all equally settled.

What is now well supported

1) AI has produced formally verified proofs for some open problems

There are documented cases where AI systems, combined with proof assistants like Lean, produced machine-checkable proofs of problems that were previously open. That is the strongest kind of evidence, because it reduces ambiguity about whether the proof is logically valid.

This includes:

• Erdős-problem results where an AI-generated proof was formalized and written up for public scrutiny.
• Larger evaluations showing AI proof-search systems can resolve a nontrivial fraction of a curated set of open Erdős problems.
• Work showing AI can help convert difficult human mathematics into formal proofs, which is itself a major advance.

2) AI has found new constructions, counterexamples, and improved bounds

Even when it does not fully “solve” a famous conjecture, AI has already produced mathematically meaningful advances:

• better bounds in combinatorics and geometry,
• novel constructions that humans had not found,
• algorithmic improvements with provable mathematical significance,
• and new pathways that help humans complete proofs.

This matters because many research breakthroughs in mathematics are not binary “problem solved” events; they are sharper bounds, new examples, or new methods.

3) There are now credible claims of AI settling long-standing open problems

The most notable examples are:

Erdős unit-distance conjecture

An OpenAI model reportedly generated a disproof of a long-standing conjecture in discrete geometry, and the result was described publicly as mathematically substantial and worthy of top-journal consideration if done by humans. This is credible evidence of a real research-level advance, though the exact role of human checking and post-processing matters for how one describes “AI solved it.”

Navier–Stokes Millennium Prize Problem

OpenAI also publicly released a proposed AI-generated solution with a Lean formalization. If correct, this would be extraordinary: it would amount to a solution of a Clay Millennium Prize Problem. However, this is still best described as a serious claim under scrutiny, not yet a universally accepted theorem. For something of this importance, independent validation by specialists is essential.

What the evidence does not yet show

Despite major progress, there is not yet broad, settled evidence that AI has autonomously solved the most famous central open problems in mathematics at the level of:

• the Riemann Hypothesis,
• P vs NP,
• Hodge conjecture,
• Birch and Swinnerton-Dyer,
• or a fully community-accepted Millennium Prize resolution across the board.

The key reason is that:

• some results are still first-party claims awaiting independent mathematical verification,
• some “solutions” may depend on formalizations that do not perfectly capture the original informal statement,
• and many successes come from human-AI collaboration rather than fully autonomous discovery.

Best current summary

The most accurate statement is:

AI has already made real, research-grade contributions to mathematics, including some formally verified resolutions or disproofs of previously open problems, and it has materially advanced many others. But only a small number of such results are currently documented at the level needed for the mathematical community to regard them as fully established.

So the answer is yes, materially advanced for sure; solved in some cases, but the set of clearly accepted, independently verified “AI solved a major open problem” examples is still limited.