Back Home

AI 研究

A Claude-Assisted Riemann Hypothesis Proof Claim Has Been Independently Formally Rechecked

Public records show that different proof checkers have accepted the formal derivation. The questions now are whether the proposition matches the problem and how mathematicians will understand the long argument, which has not yet been written up formally.

Nadar · Public domain · Image source
zh-Hant

For readers following AI’s mathematical capabilities, this case offers more concrete material to examine than a generated answer: Levent Alpöge has made public a claim of a Claude-assisted proof of the Riemann Hypothesis, accompanied by Lean formalization code and records of independent rechecking. In a statement published on November 20, 2026, he wrote that he “believe[s] [he] found a proof,” and said he still needed to understand, explain, and formally write up the argument. What can be reported at this stage is a claim supported by formal-checking records; it is not grounds to declare that the mathematical community has confirmed a breakthrough. Author’s statement, announcement date and checking records

First, confirm which problem the code proves

Lean is a proof assistant that lets people express mathematical definitions, propositions, and derivations in a precise language, with its “kernel” checking the logical steps. The first question, then, is whether the proposition in the code really corresponds to the Riemann Hypothesis.

The checking report published on Andrew Sutherland’s MIT personal webpage says the first target theorem directly uses Mathlib, the mathematical library’s RiemannHypothesis: every nontrivial zero of the Riemann zeta function has real part equal to one-half. All five target theorems have also been checked against separately written challenge propositions using con-leche’s comparison mode. This check is intended to reduce the risk that the code proves a different proposition. Proposition and comparison results

Passing checks still leaves assumptions and rechecking scope

As of November 21 at 12:55 UTC, Sutherland reported that all 175,187 modules had finished compiling and had been rechecked by the Lean kernel; nanoda and con-leche, which are separate implementations, also accepted the complete exported proof. The page notes that Sutherland performed the checks using Claude Code. These are public results from an independent checker; this article has not downloaded or run the proof code. The same page still lists some end-to-end steps as in progress, so the checks should not be described as all complete. Checking status

Another step is an axiom audit: which starting assumptions does the conclusion depend on? The public record lists the dependencies for each of the five target theorems. They rely only on Lean’s three standard foundational axioms, propext, Classical.choice, and Quot.sound, and do not use sorryAx to bypass a proof. It is important to distinguish the fact that sorryAx appears in the environment from whether these theorems depend on it; what matters is the assumptions each conclusion actually uses. Original axiom audit

Formal correctness and mathematical understanding answer different questions

The community reading project nasqret/riemann-rev says explicitly that it is not an official companion, a referee report, or an endorsement. Its account says that a complete, human-readable argument has not yet been written and that the work has not undergone traditional mathematical peer review. These points should be presented as statements from that project. Reading project and limitations

The author’s original statement directly acknowledges that, at the time of publication, he had not yet mastered much of the mathematics surrounding the argument. He decided to provide the Lean proof first, then work toward understanding and formally writing it up. This is not evidence that formal checking failed; it means that even when a checker accepts a derivation, explaining the key ideas, confirming that the definitions reflect the intended meaning, and helping other researchers understand the method remain separate tasks. Author’s statement

For people using AI to assist with research, this case offers a practical order for evaluating a claim: check the proposition, trace its assumptions, and then see whether different checkers can recheck the full derivation. As for how much Claude contributed, the main evidence currently available is the author’s account. Formal verification checks the derivation; it cannot, by itself, quantify the respective research contributions of the human and the AI.

Sources

  1. Levent Alpöge
  2. remark.pdf
  3. Verification files for the Lean proof of the Riemann Hypothesis
  4. audit.log
  5. The announced proof of the Riemann hypothesis