Can We Trust AI Math Proofs? How Lean Verifies Them

Quick answer: AI-generated math proofs can be trusted after independent verification, not because an AI says so. A proof assistant such as Lean checks a formal proof against precise logical rules and explicit assumptions. Experts must also confirm that the formal statement matches the intended theorem, that no unproved steps or extra axioms slipped in, and that essential computations have checked justification. A successful build alone is not enough evidence for an unfamiliar AI-generated proof.

Sources checked November 22, 2026.

In November 2026, a proof of the Riemann Hypothesis appeared that runs to millions of lines of Lean code. No human will ever read all of it. Within a day, the Lean and Mathlib communities certified that the statement is right and that the proof checks. Should we believe it? This question, once philosophical, is now practical. Here is how mathematicians decide whether to trust a proof produced with AI, and where the real weak points are.

Three Ways a Proof Gets Checked

MethodHow it worksStrengthWeakness
Human peer reviewExperts read the argumentChecks ideas and meaningSlow; referees can miss errors in long proofs
Formal verificationA proof assistant checks every logical stepExtremely reliable for what it checksOnly checks the statement it is given
ComputationSoftware checks cases or numbersHandles enormous case listsCode and numerics must be trusted or verified too

AI-produced proofs increasingly come with formal versions, because no human referee can keep up with machines that write proofs this fast.

How Lean Checks a Proof

Lean is a proof assistant: a programming language in which definitions, theorems and proofs are written so a computer can check them. Its key design idea, sometimes called the de Bruijn criterion, is that only a small core program, the kernel, needs to be trusted. However the proof was found, by a person, a search program or an AI, the kernel checks each step against the rules of logic. If a step does not follow, the proof fails.

That is why the author does not matter: a correct Lean proof written by an AI is as checked as one written by a Fields Medalist. It also means AI cannot “talk its way” past Lean the way it can past a tired reader. For unfamiliar, high-stakes work, reviewers still rebuild the proof independently, on their own machines, with the stated Lean and Mathlib versions. For the Riemann Hypothesis proof, MIT’s Andrew Sutherland logged several successful independent checks, with more runs still in progress on November 22.

A Valid Proof of the Wrong Claim: An Example

Suppose a headline says “every real number has a positive square.” Someone formalizes and proves, in Lean, that every nonzero real number has a positive square: if \(x \ne 0\), then \(x^2 > 0\). Lean accepts the proof, and it is correct. But it does not establish the headline, and the headline is false: \(0^2 = 0\). The checker did its job; the mismatch was in the statement. This is exactly the gap human reviewers must close.

What Lean Cannot Check

  1. Is the statement right? Lean checks that the proof proves the formal statement. If the statement was translated wrongly, for example with a missing condition, the proof is of the wrong theorem. This is why the Riemann Hypothesis certification separately confirmed that the formal statement matches the standard one.
  2. Are there hidden assumptions? A Lean file can contain unproved placeholders (marked sorry) or extra axioms. Standard foundational axioms are expected; the warning signs are unfinished steps or unusual added axioms. The command #print axioms lists every axiom a theorem depends on.
  3. Do essential computations have checked justification? Some results rely on large outside computations, which is fine if a verified checker confirms them. The November 2026 Navier–Stokes package said its numerical certificate and the computation checking it were outside the formalization, so that part needs separate checking.
  4. Do we understand it? A verified theorem can support further mathematics before anyone understands its proof, but explanation is how researchers extract methods and teach results. Terence Tao has stressed that results need explanation to become part of mathematics.

Machine Proofs Before AI: A Short History

YearResultHow trust was won
1976Four color theorem (Appel and Haken)Computer case check; formally verified in Coq by Georges Gonthier in 2005
1998Kepler conjecture (Thomas Hales)Referees were “99% certain”; the Flyspeck project completed a formal proof in 2014
2022Liquid Tensor Experiment (Peter Scholze’s theorem)Completed in Lean on July 14, 2022, checking a result Scholze had asked the community to verify
2026Fermat’s Last TheoremFully formalized in Lean with Claude’s help, reviewed by Kevin Buzzard
2026Riemann Hypothesis (Alpöge, with Claude)Lean proof; statement and proof certified by the Lean/Mathlib community on Nov 21

The pattern is old: mathematicians first doubt a computer proof, then build better ways to check it. AI speeds up the cycle.

Where AI Proof Claims Have Gone Wrong

  • Rediscoveries. Several AI “solutions” of Erdős problems turned out to be in old papers.
  • Errors in informal proofs. In a 2025 AI-assisted paper on Nesterov’s method, Jang and Ryu’s January 2026 revision withdrew an auxiliary argument after an error was found, which they attributed to insufficient checking of AI-generated reasoning; the main result stood.
  • Rushed announcements. On September 11, 2026, 25 Fields Medalists criticized results announced without write-ups, explanation or credit to prior work, a concern about practice rather than evidence that particular formal proofs failed.

6 Questions to Ask About Any AI Proof Claim

  1. Is there a formal proof (Lean, Coq, Isabelle), or only a written argument?
  2. Has someone independent checked that the formal statement matches the real theorem, and rebuilt the proof?
  3. Does the proof use any unproved placeholders or unusual axioms?
  4. Does every essential computation have a checked justification, and what software must be trusted?
  5. Have independent experts, not only the company, reviewed it?
  6. Is it new, or already in the literature?

For the 2026 results these questions apply to, see AI math breakthroughs of 2026.

Can ChatGPT Check a Proof?

Not in the same sense. A chatbot can point out suspicious steps, but its “looks correct” is an opinion, not a verification. Lean’s verdict comes from a fixed checker; a chatbot’s comes from the same kind of prediction that produced the proof. Use AI review as a first pass, then rely on formal checking or expert review. See why AI gets math wrong.

Is a Proof Nobody Read Still a Proof?

Logically, yes: a valid argument is valid whoever checks it. Socially, mathematics also asks for understanding, and that takes time. After the four color theorem, mathematicians accepted the result long before they found it satisfying. The likely future is a split: machines guarantee correctness, while people work to explain why results are true, a division of labor that mathematicians are now building rules for.

Common Misconceptions

“Lean proofs can’t be wrong.” The checking is extremely reliable, but the statement, axioms and outside computations still need human review.

“An AI proof is automatically suspect.” No. Once formally verified, who wrote it does not affect its correctness.

“Peer review is obsolete.” No. Review shifts from checking every line to checking statements, assumptions, novelty and meaning.

Watch: How Computers Check Proofs

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

Automated Mathematical Proofs (Computerphile). A clear introduction to how computers check proofs.
How to learn Lean programming language (Lex Clips (Terence Tao and Lex Fridman)). Terence Tao on why Lean matters for checking mathematics.

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. In Lean, which part of the system must be trusted?
Found a bug? Let us know!
2. What can Lean not check by itself?
Found a bug? Let us know!
3. What does sorry mean in a Lean proof?
Found a bug? Let us know!
4. Which theorem was proved with a computer in 1976?
Found a bug? Let us know!
5. Why did the Navier–Stokes package need extra checking?
Found a bug? Let us know!

Frequently Asked Questions

Can AI math proofs be trusted?

They can be trusted when properly verified: ideally a formal proof in a system like Lean, an independent check that the formal statement matches the real theorem, no unproved assumptions, and expert review. An AI’s confidence is not evidence.

How does Lean verify a proof?

Lean has a small trusted kernel that checks every logical step of a formal proof against the rules of logic. If any step does not follow, the proof is rejected, no matter who or what wrote it.

What can formal verification not check?

Whether the formal statement means the theorem people intended, whether extra axioms or placeholders were used, whether outside computations are correct, and whether anyone understands the proof.

Has a proof ever been accepted without humans reading it all?

Yes. The four color theorem (1976) relied on computer case checks, and the Kepler conjecture was formally verified by the Flyspeck project in 2014. In 2026, a millions-of-lines Lean proof of the Riemann Hypothesis was certified by the Lean community.

Is the AI Riemann Hypothesis proof verified?

Yes, formally. On November 21, 2026, the Lean and Mathlib communities certified both the formal statement and the proof released by Levent Alpöge, who developed it with Claude. That certification concerns the formal proof; a human-readable explanation and the Clay Institute’s prize process were still pending.

Can Lean prove the wrong statement?

Yes. Lean checks that a proof establishes the formal statement it was given. If the statement was written wrongly, for example with a missing condition, Lean can correctly verify a proof of the wrong claim, so people must check the statement.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Can We Trust AI Math Proofs? How Lean Verifies Them - Effortless Math"?

No one replied yet.

Leave a Reply