← Back to the latest articles
QUESTION 133AI, Mathematics & Knowledge

By NOW IN QUESTION EditorialSourcing & correctionsReport an error

Claude’s Riemann Hypothesis Proof: Lean Certification and the Question of Understanding

Formal checking, public acceptance and understanding why are different achievements

On November 21, the Lean and Mathlib communities certified the statement and proof of Levent Alpöge’s Claude-assisted Riemann result. That changes the evidence for correctness. What does it change about understanding why the result is true?

Is the problem now solved?

On November 20, 2026, a claim by Anthropic mathematician Levent Alpöge became public: he believed he had found a proof of the Riemann hypothesis while working with Claude. The four-page note on his website says the method extends to the generalized hypothesis and has been formalized without unproven assumptions. Those are the author’s claims, not findings independently certified by this publication.

Quanta’s November 20 update described compilation successes with additional verification continuing. The evidence advanced on November 21. A joint Lean–Mathlib statement, endorsed by the Lean FRO and Mathlib Initiative, certifies the statement and proof after matching Mathlib’s theorem and applying multiple checkers. Describing the result as merely its author’s assertion is now out of date.

The Guardian reported the certification on November 21. Formal correctness, human explanation and prize recognition remain distinct achievements.

Andrew Sutherland’s MIT verification record, in the retrieved version timestamped November 21 at 12:55 UTC, lists successful checks alongside some runs still in progress. That timestamped record should not be rewritten as saying every individual run had finished. Its generalized claims concern Dirichlet L-functions, not all conceivable L-functions.

How the claim reached English-language X

Quanta documented how mathematician Scott Armstrong shared a short X post with an image of the note. The post is part of its November 20 reporting. Direct retrieval of the X page was restricted, so we verified this account through the magazine’s reporting.

OfficeChai’s November 21 report also embeds the official @leanprover account’s English X announcement of the statement, dated that day. This supplies a latest-24-hour English-language discussion signal. Direct retrieval of X was restricted; we checked the embedded announcement through reporting.

This establishes English-language attention, not agreement by all English speakers, national rankings or a total post count. Attention from experts and completed expert verification are also different things.

What the hypothesis concerns

The Clay Mathematics Institute’s description states that the nontrivial zeros of the zeta function all have real part one-half. The problem is closely connected to the distribution of prime numbers. Checking many examples by computation is different from proving a statement about all of them.

Record What it establishes
Note reported publicly on November 20 Author claims a discovery with Claude and formalization, and links Lean code
Quanta’s November 20 update English X dissemination and progress in independent checking, with further checks pending
November 21 joint statement The communities certify the formal statement and proof
Our November 22 update We checked the statement and timestamped verification record; we have not independently run or audited the complete code
Clay’s problem page The retrieved page still labels the problem unsolved; that alone does not adjudicate the new argument

Public release and the author’s own research timeline are distinct. Earlier dates in a research narrative do not establish that public scrutiny or acceptance had already occurred.

What passing a Lean check means

Lean’s official validation guide distinguishes whether a theorem has a valid proof from whether its formal statement means what was intended. Definitions, dependencies and independent checkers matter for that reason.

This distinction does not diminish formal verification. Its strength is that other people can inspect procedures rather than rely on an author’s reputation or a reader’s impression. But saying only that code “passed” leaves out which statement was checked, under which assumptions and by which process.

Code length and an AI brand do not settle correctness either. We are not alleging a flaw or misconduct; we are distinguishing the stages a report needs to preserve.

Prize recognition is another process

The Millennium Prize rules require publication in a qualifying outlet, at least two subsequent years and general acceptance in the global mathematical community before consideration. A public post does not itself award the prize that day.

The absence of an award is not proof that an argument is wrong. Institutional recognition and mathematical validity are connected questions, but they are not identical.

Can we know without understanding every step?

The following is our editorial analysis. Must a person understand every line of an argument before treating it as knowledge? Such a requirement would make reliance on specialized research difficult. Yet the opposite rule—that an authority or machine’s assurance ends all questions—would abandon an important demand for explanation.

At least two values are involved. One is justified confidence that a conclusion follows from its premises. Another is understanding why it follows, and being able to carry an idea into a different problem. Achieving the first can change the starting point for seeking the second. It need not complete the second.

Evaluating AI mathematics therefore requires more than a contest between humans and machines. Who framed the question, formalized the statement, checked the dependencies and explained what remains unclear? The organization of shared knowledge matters as much as the identity of the tool.

Our Anthropic IPO article distinguishes demand for shares from evidence of public benefit. Here too, combining different evaluation procedures under one word—success—can hide what has and has not been established.

Sources and verification

Updated and checked November 22, 2026, Japan time. Added the November 21 joint statement, timestamped MIT record and coverage of that day’s English X announcement. We compared the author’s note, Lean’s documentation, Clay’s description and rules, and Quanta’s November 20 updates. We have not compiled the code or independently certified correctness, completion of all external checks or prize recognition. No news photograph or post image is reproduced.

SHARE

Share this article

Carry the question into another conversation.

DISCUSSION

Discuss this question

Write in Japanese or English. Comments are translated automatically and shared across both versions of the article.

0/2000 characters

Loading comments…