On November 20, 2026, Levent Alpöge, a mathematician at Anthropic, released a Lean file containing a proof of the Riemann Hypothesis that he says he discovered together with the company's AI, Claude. The next day, the Lean/Mathlib community issued a joint statement certifying that both the statement of the Riemann Hypothesis and its formal proof are correct. On the same day, the 20th, a draft claiming to prove the Generalized Riemann Hypothesis (GRH) without AI was also released by another mathematician, but problems were reported about eight hours later.
The Riemann Hypothesis, proposed by Bernhard Riemann in 1859, is an unsolved mathematical problem and one of the Millennium Prize Problems, for which the Clay Mathematics Institute offers a $1 million prize. The formal verification is a major step toward resolving it, but Lean certification, mathematicians' understanding and acceptance of a proof, and the Clay Institute's decision to award the prize are three separate stages.
A 17-million-line Lean proof certified the day after release
The document Alpöge published on his website is only four pages long. He believes he found the proof of the Riemann Hypothesis while working with Claude, and he says the same method can resolve the GRH and, given time, the Extended Riemann Hypothesis (ERH). However, the body of the proof he released is not an ordinary mathematical paper but Lean source code split into 14 archives.
The Riemann Hypothesis states that all nontrivial zeros of the Riemann zeta function lie on the line with real part in the complex plane (the critical line). In Lean's mathematical library Mathlib, it is formulated as follows: if the zeta function is zero at a complex number , and that point is not one of the trivial zeros , and , then the real part of is .
Alpöge explained that he did not release a human-readable paper at the same time because the current argument is extremely long and readers could not follow it as it stands. A document Claude generated on a trial basis exceeded 250,000 pages, which he has not published. He says he intends to organize the proof into a form humans can understand.
He also described Claude's role concretely. He set the initial direction of combining two ideas and framed the problem, but by the later stages of the research Claude's progress had far outstripped his own. The idea of making the length of the chain of theta lifts used in the proof was also Claude's suggestion. On the other hand, he admitted that parts of the proof, such as automorphic forms on high-rank groups, lie outside his expertise, and that some of the explanations Claude generated amount to low-quality AI summaries.
The Lean/Mathlib community published a joint statement at 13:55 UTC on November 21. Backed by Lean FRO and the Mathlib Initiative, the statement says explicitly that it certifies both the statement and the proof of the Riemann Hypothesis as correct.
According to the joint statement, the verification confirmed the following four points:
- The statement being proved matches Mathlib's existing definition
RiemannHypothesis. - Dozens of people independently reproduced the compilation of the code.
comparatorandcon-leche --compareconfirmed that the proved statement had not been swapped for another.- Multiple independent proof checkers, including Lean's official kernel, accepted the proof.
The published Lean library runs to about 17 million lines. According to the verification record published by Andrew Sutherland of MIT, the files are split into 175,187 modules. This is on a scale incomparable to an ordinary mathematical paper, and checking the proof line by line by hand is not realistic.
The scope of the certification needs care, however. Sutherland listed five statements as verification targets: the Riemann Hypothesis, its equivalent formulations, and three statements about Dirichlet L-functions. The Dirichlet L-function statements cover, among other things, the absence of zeros in the region where the real part exceeds and the location of critical-strip zeros on the critical line.
Sutherland states clearly that the GRH here refers to Dirichlet L-functions for all moduli and Dirichlet characters, and does not extend to Dedekind zeta functions, Hecke L-functions, or general automorphic L-functions. What the joint statement explicitly certified is also the Riemann Hypothesis as formulated in Mathlib, and the prospect of the ERH that Alpöge mentioned is outside the scope of this formal verification.
Alpöge had also reported partial results on the Riemann Hypothesis before this announcement. On August 13, he and Ralph Furman posted to arXiv a paper showing that at least two-thirds of the nontrivial zeros of the Riemann zeta function are simple and lie on the critical line. The previous unconditional lower bound was , and they say refinements of the method can raise it to 0.6725. In September, Youness Lamzouri of the University of Lorraine also reported a result exceeding 67.25% by a different, non-AI method. If the present claim is correct, the Riemann Hypothesis would hold for all nontrivial zeros, rather than for a lower-bound proportion of them.
What does Lean certification guarantee?
Lean is a proof assistant in which mathematical statements and proofs are written in a formal language and their logical correctness is checked by computer. When a proof program is checked, a relatively small verification program called the kernel confirms that each inference follows the rules.
However, Lean's official documentation, "Validating a Lean Proof," explains that while it can confirm that a statement has been proved, it cannot guarantee that the statement is the one a human intended. If a statement different from the original problem is set up, then even a correct proof of that statement does not mean the original problem has been solved.
Lean provides mechanisms to detect such problems and improper proofs. #print axioms displays the axioms a proof depends on and checks whether anything beyond the three standard axioms, propext, Classical.choice, and Quot.sound, or sorry, which allows unproved parts, is used. lean4checker re-verifies compiled files. Further, comparator can compare a statement written separately in a trusted environment against the statement addressed by the submitted proof.
Such strict verification is needed because problems that slip past ordinary checks have actually been reported. On September 9, the security firm Trail of Bits published a way to make a fake "proof" of Fermat's Last Theorem go through in Lean. It did not exploit Lean's kernel itself but the computation feature native_decide, creating a problem where the compiler's behavior also has to be trusted.
Before this certification, Kevin Buzzard of Imperial College London had also laid out on November 12 four conditions for evaluating AI-generated Lean proofs. The items confirmed in the joint statement correspond to these. At the same time, he pointed out that formal verification in Lean does not mean that a human-oriented paper said to correspond to it is also correct.
What MIT's additional verification confirmed
In Sutherland's verification record, as of 12:55 UTC on November 21, about an hour before the joint statement, all 175,187 modules had compiled without errors, and it was confirmed that the five main theorems depend only on Lean's three standard axioms.
The independent checkers nanoda and con-leche also accepted the proof, with the latter verifying 1,468,418 declarations. lean4lean also accepted all modules, but five of them hit the default limit on computation steps, so they were rechecked with the limit raised.
As of 12:55, however, the full lake comparator run and the re-check of the entire exported proof with Lean's standard kernel and the paranoid kernel were still in progress. The checks on statement matching and axioms had passed, but not every additional verification step had been completed.
The computing resources needed for verification are also enormous. The published source archive is about 239 MB, but the proof exported for verification expands to about 61.6 GB. The procedure for running comparator reportedly requires about 100 GB of memory. On Sutherland's page as of November 22, no updates to the verification results after 12:55 could be found.
Mathematicians' understanding will take more time
Even before the Lean/Mathlib certification, some mathematicians had said it would take considerable time to understand this proof.
On November 20, Terence Tao wrote on his blog that the proof is a very large "Lean-only proof" and that it will take time to understand its overall picture. Initial analysis suggests it makes broad use of modern results from the Langlands program and arithmetic geometry.
James Maynard of the University of Oxford also said that while he expects it to be confirmed correct in the end, even checking the Lean verification results will take time, and understanding the mathematical content will take considerably more effort. Lamzouri said it could take months to years, and Tristan Buckmaster expressed a similar view to Nature. Meanwhile, Andrew Granville of the University of Montreal sharply criticized the content of the published documents, so mathematicians' assessments are not yet unified.
What mathematicians seek from a proof is not just a guarantee of correctness. It is also important to understand which ideas broke through the hard parts and whether the method can be applied to other open problems. Since Alpöge himself admits he does not fully understand parts of the proof, organizing the mathematical argument out of 17 million lines of code will likely require the cooperation of many experts.
There are precedents in which even major proofs took a long time to be accepted by the mathematical community after announcement. Andrew Wiles announced his proof of Fermat's Last Theorem on June 23, 1993, but a problem was later found, and the corrected paper written with Richard Taylor was published in the Annals of Mathematics in 1995.
Among computer-assisted proofs, Thomas Hales's proof of the Kepler conjecture is a representative example. The paper submitted on September 4, 1998 was accepted on August 16, 2005, taking about seven years. The subsequent formalization project Flyspeck, using Isabelle and HOL Light, was completed on August 10, 2014, about 15.9 years after the first submission. The formalization of Peter Scholze's "Liquid Tensor Experiment," posed on December 5, 2020, also took 586 days until completion on July 14, 2022.
A proof of the GRH without AI: problems reported eight hours after release
On the same day as Alpöge's announcement, November 20, a proof of the Generalized Riemann Hypothesis by another mathematician was also released.
Ravi Vakil, president of the American Mathematical Society (AMS) and a mathematician at Stanford University, published a blog post titled "Hurried Announcement: Alessandro Masullo GRH" at 18:18 UTC on November 20. It reports that Alessandro Maria Masullo had produced a proof of the Generalized Riemann Hypothesis.
Masullo is a mathematician who earned his PhD at Stanford University under Brian Conrad, and Conrad had examined the draft in detail over several months. According to Vakil, Conrad initially judged that the argument appeared to be working. The draft was reportedly written entirely by humans, without AI.
The original plan was to spend about another year polishing the text and having a sufficient number of experts check it before announcing. But after Alpöge's release of the Lean proof, they decided to publish the draft in its current state.
However, about eight hours after publication, Vakil added an update to the blog post. Several mathematicians had reported problems they discovered with the help of AI models. Vakil has not given a judgment on whether the basic approach of the proof is correct.
Conrad also told Scientific American that the draft is in the middle of substantial revision and that whether the method succeeds cannot be judged until the paper is closer to completion. Masullo's proof differs from Alpöge's in that it had been reviewed by experts before release, but problems were reported after publication and revision work is now ongoing. No record of completed formal verification, in Lean or otherwise, can be found.
The $1 million prize requires two years of scrutiny after publication
Lean/Mathlib certification does not mean the $1 million prize set for the Riemann Hypothesis will be awarded immediately.
The Clay Mathematics Institute adopted revised rules for the Millennium Prize Problems on September 26, 2018. According to the official rules, three conditions must be met before a proposed solution is considered for review:
- The proposed solution has been published in a Qualifying Outlet.
- At least two years have passed since that publication.
- The solution has received general acceptance in the global mathematical community.
Furthermore, the PDF version of the rules also requires that the proposed solution adequately address the official problem statement. Whether this is satisfied is determined at the Clay Institute's sole discretion.
A Qualifying Outlet refers to internationally respected peer-reviewed mathematics journals and the like. The rules set specific requirements concerning editorial boards and review procedures, and simply posting to arXiv or GitHub does not satisfy them. However, Section 6(f) also contains a provision allowing the Institute, at its discretion, to relax or waive the publication requirement if it receives expert advice that the proposed solution is likely correct.
Even after the two-year waiting period ends, review does not begin automatically. The Clay Institute judges whether the solution has been sufficiently vetted by the mathematical community, drawing on papers and books by independent researchers, discussions at international conferences, and so on. If it finds the solution worth examining in detail, it convenes a Special Advisory Committee of at least three members, at least two of whom are experts in the relevant field. The Scientific Advisory Board considers the committee's report, and the Board of Directors makes the final decision on the award.
The only case so far in which a Millennium Prize was actually awarded is Grigori Perelman's solution of the Poincaré conjecture. Perelman posted his first preprint to arXiv on November 11, 2002, and released two more papers in 2003, but he did not submit them to peer-reviewed journals.
As a result, detailed expositions and papers by other mathematicians became the basis of verification. In August 2007, the exposition Ricci Flow and the Poincaré Conjecture by John Morgan and Gang Tian was published, and in 2008 an exposition by Bruce Kleiner and John Lott was published in Geometry & Topology. A January 2009 article in Communications of the ACM reported that the two-year waiting period began from the expositions published in 2007.
The Clay Mathematics Institute announced the award of the prize to Perelman on March 18, 2010, 2,684 days after the first preprint was released. Perelman, however, declined the $1 million prize. Moreover, the award was made before the 2018 rule revision, so it is not an example showing that the current rules would be applied the same way as then.
Where things stand across four hurdles
Comparing Alpöge's and Masullo's claims across four stages (formal verification, understanding by mathematicians, peer review and publication, and the Clay Institute's rules), the current situation is as follows.
| Stage | Alpöge's claim (RH and GRH) | Masullo's claim (GRH) |
|---|---|---|
| Formal verification in Lean | Statement and proof of the Riemann Hypothesis certified on November 21. The Dirichlet L-function version of the GRH is also within the verification scope. Some additional verification is in progress | No record of completed formal verification can be found |
| Understanding and confirmation by experts | In progress. Tao and others have carried out initial analysis; full understanding may take months to years | In progress. Conrad examined it for months, but problems were reported after release and it is being substantially revised |
| Peer review and publication | Not done. No human-readable proof paper has been released | Not done. Draft released only on a blog |
| Clay rules: two years after publication | Clock not started | Clock not started |
Based on information confirmed as of November 22, 2026.
At present, the Clay Institute's two-year waiting period has not begun for either claim. Even if a paper were published in a qualifying outlet on November 22, 2026, the two-year condition would not be met until after November 22, 2028. In practice, the paper must first be written and peer reviewed, and after publication it must also be vetted and accepted by the mathematical community.
Going forward, a key question is whether Alpöge can organize the 17-million-line Lean proof into a human-readable paper and have it published in a peer-reviewed outlet. The results of the full lake comparator run and the additional verification with the standard and paranoid kernels that Sutherland is carrying out also need to be checked. In parallel, if Masullo's revised draft, which Conrad continues to examine, is released, evaluation of the validity of the human-written proof will also move forward.
