# Claude and the Riemann Hypothesis: What Lean Actually Verified

Independent checkers accepted Levent Alpöge’s formal proof. Here’s what they checked, and how that differs from Clay’s official status.

By Paras Madan | 2026-11-22


On November 20, mathematician Levent Alpöge released a Lean proof of the Riemann Hypothesis developed with Claude. The Lean and Mathlib communities have since certified that the proof is valid and that it proves their standard statement of the hypothesis. The Clay Mathematics Institute still lists the problem as unsolved.

## What Claude helped prove

The Riemann Hypothesis concerns the zeros of the zeta function. It says every nontrivial zero has real part one-half. That tells mathematicians how tightly they can bound the difference between the number of primes up to a given point and the expected count. Checking even trillions of individual zeros cannot settle a claim about all of them.

In [his four-page announcement](https://alpo.ge/remark.pdf), Alpöge describes starting from a simpler mathematical setting and guiding the approach towards the original problem. He credits Claude with a key choice about how many times to chain a construction called a theta lift, and with helping formalize the argument. He also claims the generalized result for Dirichlet L-functions. An extension beyond that is something he hopes to prove later.

## What Lean actually checked

Lean is a programming language and proof assistant. Its kernel checks whether a formal proof follows from the definitions and axioms in use. If Claude suggests a plausible step that fails to follow from the previous steps, the checker requires a valid justification for it.

The [Lean and Mathlib statement](https://leanprover-community.github.io/blog/posts/riemann-hypothesis/) describes two checks that are easy to lump together. First, reviewers put the statement to be proved in a separate file that imports only Mathlib, Lean’s mathematical library. Two independently implemented tools compared that target with the theorem in Alpöge’s proof. Second, several proof checkers, including Lean’s kernel, accepted the exported proof itself.

**A proof can pass a checker and still answer the wrong question if the target was defined incorrectly.** That is why it mattered that mathematicians had already vetted Mathlib’s statement, and why reviewers compared it with the submitted theorem instead of trusting the theorem name.

## How to read the verification for yourself

Start with the [community verification repository](https://github.com/leanprover-community/riemann-hypothesis). The `src/Challenge.lean` file contains the independently specified targets; `export-targets.txt` shows which declarations are pulled from the submitted proof. Those files are a better starting point than downloading the entire archive and searching it for a theorem named `riemannHypothesis`.

Then read [Andrew Sutherland’s verification log](https://math.mit.edu/~drew/grh/), especially the rows for statement comparison and axiom checks. Sutherland reports that all five headline theorems depend only on three standard Lean axioms, and that an independently implemented comparator matched all five against the challenge statements. Read each row on its own: that result does not depend on every separate checker run in the log having finished.

## What Clay’s unsolved status means

The [Clay Mathematics Institute’s problem page](https://www.claymath.org/millennium/riemann-hypothesis/) still marks the hypothesis as unsolved. Its linked prize rules require a qualifying publication, at least two years after publication, and general acceptance in the mathematics community before Clay will consider a proposed solution. That process is separate from the checks the Lean community has already run.

There is also a difference between checking a proof and understanding why it works. Alpöge released the argument as Lean code and says he plans to write it up. A readable explanation could make it much easier for another number theorist to see which parts of the argument apply to problems other than this one.

An AI-assisted result can now arrive as a proof other people can independently check before anyone has a good way to teach it. The next thing to look for is which ideas survive when mathematicians try to make the argument shorter.

I’m Paras. I help AI founders explain technical work in a way developers can inspect and understand.

[Book a 30-minute call](https://cal.com/paras-madan/30min)
