The proof of the Riemann Hypothesis aided by Claude has passed machine verification—what else do humans need to check?

popularization of scienceNovember 22, 2026release AiFun
44 0

A proof consisting of approximately 17 million lines of code—machines can verify the reasoning within it, but first they must confirm that it proves the original problem. On November 20, mathematician Levent Alpöge He published the proof of the Riemann Hypothesis that he developed in collaboration with Claude....provides a formalized Lean codebase rather than a mathematical paper written in the conventional style. On November 21,The Lean and Mathlib Communities Issue a Joint Statement, confirming that both the formal statement and proof of the classic Riemann Hypothesis have been verified. The statement is endorsed by Lean FRO and the Mathlib Initiative.

First, check the problem: Just because you’ve calculated many zero points correctly doesn’t mean you’ve proven them all.

The Classic Riemann HypothesisConsider the non-trivial zeros of the Riemann zeta function: Are the real parts of all these zeros equal to 1/2? You can visualize this as a vertical line in the complex plane; the conjecture states that all non-trivial zeros lie on this line. This question is important because the locations of the zeros are closely related to the extent to which the distribution of prime numbers deviates from the mean.

Here’s a verification question that’s easy to overlook: Suppose someone submits code that proves “within a given finite range, all verified zeros lie on the line.” Even if the proof is correct, it does not answer the question, “Without specifying a range, do all non-trivial zeros lie on the line?” This example is not meant as a criticism of this particular proof, but rather to illustrate why the review process cannot jump directly from “the code compiles” to “the conjecture has been proven.”

During the actual verification, the community created a separate file that relied solely on the Mathlib library, in which they wrote out the standard propositions to be proven, and then used Comparator cap (a poem) con-leche Comparison and verification. Both tools confirmed that Alpöge’s code proves the same Riemann Hypothesis as defined in Mathlib. Thus, the author of the proof did not specify which version to use for verification.Verification Notes for the Joint Statement

Re-verify the reasoning: It must be reproducible on other machines and must pass through different checkers.

After verifying the problem statement, the given reasoning must also be validated. The joint statement noted that dozens of participants reported successfully compiling the codebase and reproducing the same original proof derivation. Multiple verifiers then validated the exported content, including Lean’s official kernel and third-party implementations. Here, the term “kernel” refers to the core program that checks whether formal reasoning complies with the rules. Having different implementations verify the same proof is more convincing than seeing a single successful compilation: it reduces the risk of relying entirely on a single implementation.

Published by MIT mathematician Andrew SutherlandIndependent Verification RecordLists the results, logs, and checksums. For example, a standalone Rust kernel nanoda A complete proof export was accepted; the record also accurately noted the checks that were still running at the time. Sutherland’s record also verified the generalized version involving Dirichlet L-functions, but the joint statement explicitly certifies thatThe Classic Riemann Hypothesis...this should not be used as a basis for writing all more general L-functions as having already been verified.

What other work remains to be done regarding readable mathematical explanations and award deliberations?

Machine verification has yielded positive results regarding formal reasoning. Next, mathematicians face a different task: organizing the vast amount of code into an argument that peers can read, discuss, and reuse. For example, which axioms truly underpin the key reasoning? Can the tools used be simplified? What other similar problems can this approach address? A readable explanation does not necessarily require a line-by-line translation of Lean; what is more important is to clearly articulate the structure and key ideas. The joint statement also explicitly expresses the expectation that the mathematical community will next explain the new ideas presented in this proof.

As of the verification on November 22,Cray Mathematical Institute Problem PageIt is still marked as “Unresolved.” This reflects the status on the organization’s website and in the awards process; it does not constitute a rejection of the Lean community’s technical validation. According to the institute’s announcement,Rules for the Millennium Grand Prize Review...A proposed solution must be published through an approved channel, have been published for at least two years, and be widely accepted by the global mathematical community before it will be considered for review. The date the code was released cannot be directly considered the date of valid publication, and there is no basis for claiming that the prize has already been awarded.

This case also offers valuable insights for other AI-assisted research. When preparing to publish a result, first save the proposition to be proven separately, then provide evidence that can be verified by external parties, and finally clearly state the scope of the result. In the Alpöge case, what truly enabled the outside world to evaluate the results were the publicly available formal proofs, independently defined objectives, and verifiable results that could be cross-checked.

© Copyright notes

Related articles

No comments

none
No comments...