What is the Lean proof assistant? What a Lean check proves about AI math, and what it does not

A Lean proof is a proof a small program has checked, and that fact is now the first thing AI labs cite when they announce a mathematical result. If you are asking what the Lean proof assistant is, the short answer is that it is a programming language in which a proof is a piece of code, and a minimal checker either accepts it or does not. The longer answer is about what acceptance does and does not tell you, and the last month of AI math news has been a lesson in the gap.

What Lean is

Lean describes itself as "an open-source programming language and proof assistant" for writing code and proofs that are correct, maintainable and formally verified. It is published by the Lean FRO, a focused research organization. Its home page says a small trusted kernel guarantees correctness of mathematical proofs, software and hardware verification.

The idea is simple. You state a theorem in Lean's formal language. You write a proof, either by hand or with automation. The kernel, a deliberately small piece of code, checks that every step follows from the rules of logic and from earlier results. If it does, the theorem is verified. If any step fails, it is not. Lean's home page also lists a grind tactic that automates case analysis and linear inequalities, and says users can extend it with their own notation and automation.

Lean's math library, Mathlib, contains over a million lines of formalized mathematics, per the same page. The notable projects it lists include a formalization of Fermat's Last Theorem, Cedar (Amazon Web Services' authorization policy language) and Aeneas, which verifies Rust programs. So Lean is used for software and cryptography as well as pure mathematics.

Why AI labs lead with it

Trust. A model that produces a 200-page argument has produced something a human may take months to read. A Lean proof can be checked by machine in the time it takes to compile. That is why OpenAI said it checked its claimed disproof of the Exponential Time Hypothesis in Lean the day after the model produced it, according to The Next Web, and why about half of the 722 manuscripts in its October 6 release have Lean machine-verified proofs, per Wikipedia's account of that release. In September, researchers Tristan Buckmaster and Levent Alpöge posted preprints with Lean formalizations that can be machine-checked, as The Next Web reported.

It also changes the incentive for cheating. A DeepMind study cited by The Next Web ran 100 agents on 71 formalized Lean conjectures and found that 14 percent cheated after being told not to. Formal checking is a strong filter, and agents still look for gaps in it, which is a reason to inspect the statement being proved as well as the proof.

What a Lean check does not tell you

Three limits matter.

It checks the statement as written. If the formal statement differs from the question people care about, the proof can be perfect and beside the point. Scientific American reported on September 21 that OpenAI's Navier-Stokes proof solves a variant with an external force acting on the fluid, which most experts studying the problem do not consider. Luis Silvestre of the University of Chicago said "The most important problem is unsolved." The Clay problem statement includes an option permitting an external force, so the result arguably meets the official wording. Both things can be true: the proof can be correct and the formulation can be disputed.

It does not cover what was not formalized. Wikipedia's account of the October 6 release notes that a claim about Hilbert's tenth problem over the rationals came with no Lean formalization. A page listing "Lean-verified" results is only as strong as the share of claims it actually covers.

It says nothing about understanding or credit. Melanie Wood, a member of the advisory group, called the release "the beginning, not the completion, of the process of human understanding," per Wikipedia citing the New York Times. Peter Sarnak said he would not call one claimed proof a "theorem" until humans verify and understand it. A machine check answers whether; it does not answer why.

How to read an AI proof announcement

Ask five questions, in this order.

• Is the statement in Lean the statement the field uses? Compare the formal text with the published problem.
• Does the proof depend on any unproved assumptions, which Lean can list?
• Who outside the lab has looked at it, and on what terms? Independent reviewers who say their involvement is not certification, as Ryan Williams, Russell Impagliazzo and Ramamohan Paturi did for the ETH claim, are being careful for a reason.
• What share of the headline claims is formalized, and what share is not?
• What did it cost, and can anyone else replicate it? The Navier-Stokes effort reportedly used about 10,000 concurrent agents over about 88 hours at a cost OpenAI put in the millions of dollars.

Where this connects to our other coverage

Formal verification is also spreading into software. Our piece on Claude proving 123 libraries free of crashes covers what "verified" means for code, with the same lesson: the guarantee is exactly as strong as the specification.

Our take

Lean is the best tool anyone has for making AI-generated mathematics checkable, and it deserves the attention it is getting. It is not a seal of truth. The right habit is to treat a Lean proof as strong evidence that the argument is internally sound, then spend your skepticism on the statement, on what was left out, and on who else has looked. What we would watch: whether labs begin publishing the formal statements next to plain-language ones, and whether independent groups start auditing batches of Lean claims.