Executive Summary
A mathematician at Anthropic says he has found a proof of the Riemann Hypothesis. On November 20, Levent Alpöge posted a four-page note and released the body of the proof as 175,187 files of Lean 4 code, a language for writing mathematical statements so a computer can check them one step at a time. Andrew Sutherland of MIT and other outside mathematicians ran the code themselves and confirmed that it compiles with no errors.
The machine letting it through does not mean a person has read it. Alpöge did not publish a human-readable account alongside the code. Written out in ordinary prose, he said, it would run past 250,000 pages, and he could not vouch for what came out. The objection Kevin Buzzard of Imperial College London raised was not about the code being wrong either. It was about whether the statement written inside the code is the statement we call the Riemann Hypothesis. A day later, after the verification groups said the proof is correct, Alpöge wrote that he does not know why the theorem is true.
Sections 1 through 4 set out what public materials and news coverage establish. Section 5, which reads the episode as the distance between a verifiable output and a trustworthy one, is this article's own interpretation.
Key Numbers
Sources: Scientific American (2026-11-20) · the same outlet's follow-up (2026-11-21) · Clay Mathematics Institute
The first two numbers show how large this proof is from a human point of view. The third marks how far the machine side of the checking has gone, and the fourth is the time the institutions of mathematics require before any of this counts as settled.
175,187
Lean source files released
Around 17 million lines, 239.3 MB compressed
250,000 pages
Length if written out for human readers
Alpöge's own estimate, and his reason for not writing one
3 axioms
Assumptions the core theorems rest on
Sutherland confirmed nothing hidden beyond Lean's three standard axioms
2 years minimum
Public scrutiny required for a Millennium Prize
Peer-reviewed publication comes first, and that clock has not started
A Four-Page Note and 175,187 Lean Files
What Alpöge put on his own website on November 20 was four pages long. It says he believes he has found a proof of the Riemann Hypothesis together with Claude, and the proof itself is not in it. What came with it instead were 175,187 source files written in Lean 4. They take up 239.3 MB compressed, more than 61.5 GB unpacked, and run to roughly 17 million lines. He discussed the release with Anthropic chief executive Dario Amodei beforehand, according to reporting.
The missing human-readable account was not an oversight. Alpöge said that writing this proof out in ordinary mathematical prose would take more than 250,000 pages, and that what came out of that exercise would not be worth reading. So the only published form of the proof is code that machines read.
The Riemann Hypothesis is a conjecture Bernhard Riemann arrived at in 1859 while trying to describe how prime numbers are distributed. It says that every non-trivial zero of the Riemann zeta function sits on a single line, the one where the real part is 1/2. If that holds, we can say far more precisely than we can today how primes scatter along the number line. For 167 years nobody has shown it either true or false, and in the meantime hundreds of theorems have been built on top of it by assuming it. This is the statement Alpöge says he has proved.
The scope is not the Riemann Hypothesis alone. Alpöge claims the work also covers the broader Generalized Riemann Hypothesis, which is a claim not about the zeta function on its own but about a whole family of functions defined the same way. The last piece reportedly fell into place on the night of November 15. He has described Claude as handling algebraic manipulation, trace formulas and technical lemmas, and as blowing past him entirely near the finish line. Which model was used has not been disclosed.
The response came within a day. Andrew Sutherland of MIT took the code and recompiled it independently, confirming that all 175,187 modules go through without errors. He also confirmed that five core theorems assume nothing beyond Lean's three standard axioms. Sutherland called it a historic moment.
What the Compiler Actually Passed
In a proof assistant like Lean, saying that code goes through has a precise meaning. It means there is no logical gap between the premises written in the code and the conclusion written in the code. The kind of mistake a reader senses when a paper slides from one line to the next without earning it is caught entirely at this stage. That is why mathematicians trust the check.
The point where Kevin Buzzard applied the brakes lies outside that. He said that when he judges a document with Lean code attached, there is a first question he asks: whether the statement of the theorem in the code matches the statement of the actual theorem. Code going through means something logically coherent has been proved, and nothing in that result guarantees the something is the Riemann Hypothesis.
This is not an abstract worry. Formalization is human work. If the definition of a single symbol shifts slightly while a mathematical statement is translated into Lean syntax, the machine will faithfully prove the shifted statement and faithfully let it through. The fact that what passed differs from what someone meant to ask is recorded nowhere in the code as an error.
The procedure Buzzard asked for is called a comparator run. It checks two things. One is whether the code quietly redefines the meaning of symbols used in the statement. The other is whether the same result comes back when the proof is recompiled in paranoid mode across several independent kernels. The second check exists to catch a proof that got through by exploiting a still unknown bug in Lean's official kernel.
This is the point where being verifiable and being trustworthy part company. A machine confirms almost perfectly whether a proof holds up inside its own premises, and whether those premises are a faithful translation of the question we had falls outside what it checks. The pass is on the record. What passed is still something a person has to go and read.
What Remains After the Comparator Run
The comparator run was done within a day. Around 09:30 UTC on November 21, word circulated that the statement comparison and the axiom check had passed and the kernel check was still running. At 13:55 the same day, the Lean FRO and the Mathlib Initiative issued a joint statement saying they are entirely confident that Alpöge's proof of the Riemann Hypothesis is correct. The Lean FRO builds the Lean language and its kernel; the Mathlib Initiative curates the library of mathematics written in Lean. The people who built the tools this proof depends on are also the people who ran the check.
A line carried inside that same statement changes the character of the episode. Alpöge wrote that at this moment we do not even know why the theorem is true. He added that he does not yet know how much of this mathematics is a new idea and how much is standard work the machine carried out. The person who produced the proof cannot explain what is in it.
The mathematician James Maynard put his finger on exactly that gap. He said that for now this means almost nothing to him, because he never doubted that the Riemann Hypothesis is true. What the field wanted from this problem was not a verdict of true or false but insight into why it is true, and the code that has been published has not handed that insight to anyone yet.
There is institutional time to run as well. The Clay Mathematics Institute's rules for a Millennium Prize require publication in a peer-reviewed journal, followed by at least two years of public scrutiny. Between the moment a machine says a proof is correct and the moment mathematics accepts it there is a formalized distance, and in this case the clock has not started.
The Same Day, an All-Human Draft Appeared
Hours after Alpöge's announcement, another draft proof of the Riemann Hypothesis came out of Stanford. According to the-decoder, it was written by Alessandro Masullo, who had planned to spend a year polishing it and moved the schedule forward. Brian Conrad gave feedback over several months, and Professor Ravi Vakil said the work is purely human, with no AI involved.
What followed is awkwardly fitting. Within eight hours of publication there were reports that problems had been found in this human proof with help from several AI models. Over a single weekend, AI stood in the prover's chair on one side and the reviewer's chair on the other.
That the two proofs arrived on the same day upends the impression a one-line headline leaves. News that a machine has proved something reads easily as news that there is less for people to do, but what has actually grown since November 20 is the verification labor that falls to humans. Comparing statements, running kernels repeatedly, cross-reading a draft, asking again why the thing is true: all of it has piled up on the human side.
Why Pebblous Is Watching This Case
Read as an internal affair of mathematics, this episode offers nothing to take back to data work. Strip it down to its structure and the picture changes. An automated check announced a pass, and what that check had actually confirmed could only be learned by a person reading it separately. This structure repeats daily outside mathematics.
The warning was already in the literature. An arXiv paper released in 2026 used the Navier-Stokes case to set out the point that a proof produced by autoformalization can go through in Lean without that establishing it has proved the original natural-language theorem. What formal verification guarantees is narrower than what people expect it to guarantee.
In data and model work the same structure shows up in plainer clothes. The tests all came back green, the benchmark score cleared the bar, the API response matched the schema. All three are true. Whether that test reproduces the failure we were worried about, whether that benchmark carries the distribution we actually serve, whether that schema carries the meaning we intend to use: none of it is contained in the pass signal. A pass is an answer to the question the check defines, not an answer to the question we had.
What mathematicians did this week was not an exotic procedure either. All of it comes down to what they read after receiving the record of a pass. For anyone receiving AI output, the three things to confirm sit in the same place.
- Is it written down somewhere what question this check asks? If the name of the check survives and the definition does not, a pass record guarantees nothing.
- Has a person read the question used in the check and compared it against the question we actually want answered? This is Buzzard's first question.
- After the pass, is the result all that remains, or does an account of why that result came out remain as well? Without the account, the next similar output arrives with nothing to judge it by again.
If AI-Ready Data is understood only as cleanly refined values, this is exactly the spot that empties out. What a value means, and on what basis it was recorded that way, has to travel with the data for a person to be able to re-read a machine's verdict. With verdicts and no meanings, pass signals accumulate while the people who can interpret them thin out.
Whichever way the Riemann Hypothesis story lands, one thing was settled this week. A proof a machine will pass and a proof a human can read have visibly come apart for the first time. Closing the gap is still human work, and who carries that work, under what procedure, is the question ahead.
Editor's Note
Alpöge's four-page note and the Lean repository could not be opened directly, so the file count of 175,187, the sizes, the 250,000-page estimate, Sutherland's recompilation result and Buzzard's remarks all come through Scientific American and its follow-up coverage. The line about the release being discussed with Amodei beforehand rests on the same reporting. The 1859 date and the description of the zeta function's zeros are textbook background and are not separately sourced. The 13:55 UTC joint statement of November 21, the Alpöge remarks quoted inside it, and Maynard's reaction come by the same route. Remarks by Alpöge, Sutherland, Buzzard, Vakil and Maynard are given here as indirect speech drawn from that coverage rather than as verbatim quotation. The author's name for the competing Stanford draft and the report of problems found within eight hours rest on the-decoder and could not be confirmed at first hand. The Clay Mathematics Institute's two-year rule comes from the institute's published rules. For the arXiv paper on autoformalization, only the abstract was checked. The reading in Section 5 is this article's own judgment and is not in the public materials. This is a fast-moving story, and this article carries only what was confirmed as of November 22, 2026.
References
News Coverage
- 1.Scientific American. (2026). "Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math." The initial announcement and Buzzard's objection.
- 2.Scientific American. (2026). "The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest." The comparator run result and the joint statement.
- 3.the-decoder. (2026). "Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft." The competing Stanford draft.
- 4.Al Jazeera. (2026). "Riemann hypothesis: Has the world's most famous maths problem been cracked?."
- 5.Etoday. (2026). "Has AI solved a 167-year-old mathematical problem? Anthropic researcher claims a Claude proof of the Riemann hypothesis." Korean-language coverage.
- 6.YTN. (2026). "Claim that AI has proved the Riemann hypothesis, a 167-year-old problem, stirs the field." Korean-language coverage.
Academic and Official Documents
- 7.arXiv:2610.08144. (2026). "Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs." The gap between formal verification and a natural-language theorem.
- 8.Clay Mathematics Institute. "Millennium Prize Problems." Peer-reviewed publication and the minimum two years of public scrutiny.