Claude Proved Riemann Hypothesis? Hold the Applause
Anthropic researcher claims Claude proved the Riemann Hypothesis, but top mathematicians urge extreme caution before celebrating. The timing—days after Anthropic’s IPO—only deepens the skepticism.
A proof, but can anyone read it?
Levent Alpoge, a mathematician who also works at Anthropic, posted a four-page claim on his personal website on November 20: Claude, the company’s large language model, has proved the Riemann hypothesis. He also wrote a formal verification in Lean, the programming language used to mechanically check proofs, and said the work spanned more than 250,000 pages of interaction with the AI.
He then added the detail that matters most for anyone evaluating the claim honestly. “Near the finish line,” Alpoge wrote, Claude “flew way ahead and danced and sparkled,” and left him behind. He does not fully understand the argument himself.
That sentence is not humility. It is a warning label.
Why the math community should remain skeptical
The reaction from established researchers was immediate and measured. Brian Conrey, director of the American Institute of Mathematics and one of the world’s leading authorities on the Riemann hypothesis, told AFP he had read only the website post, not the manuscript itself. He acknowledged the potential magnitude of the result—calling it possibly “the biggest event in the history of mathematics” if true—but followed with a crucial qualifier.
“It seems quite complicated,” Conrey said. “I believe there will be much simpler proofs.”
In practice, when someone claims a proof that is extraordinarily complex and no one else can yet follow, the usual next step is a simpler proof. That pattern has held for centuries. The real test is whether independent experts can reconstruct every step, and so far none have done that for Alpoge’s claim.
Kit Yates, a professor at Bath University who studies mathematical biology and science communication, acknowledged the stakes. If the result survives peer review and two years of community examination—the Clay Mathematics Institute’s standard for confirming a Millennium Prize solution—it would reshape understanding of both mathematics and AI capabilities. But he emphasized the same point: this needs “enteratic verification.”
Nestor Guillen, a visiting professor at New York University, offered the sharpest take. He called Alpoge’s candid admission that he does not fully grasp the proof a positive sign—not a flaw in character but evidence about the nature of AI collaboration. The process described was not a machine deity handing down truth. It was a tool generating suggestions that still require human judgment, and human judgment admitted uncertainty at the critical juncture.
The OpenAI shadow
This claim cannot be read in isolation. In September, Alpoge publicly accused OpenAI of using research data from his prior collaboration with NYU’s Tristan Buckmaster on the Navier-Stokes millennium problem without attribution. Sam Altman dismissed the allegation as baseless plagiarism claims. The episode left the mathematics community deeply uneasy about how AI-generated results intersect with academic ownership and rigor.
Then, last month, OpenAI released hundreds of papers making various claims about solving difficult problems. Many were met by experts with a mix of admiration and alarm: the output was impressive, but it mixed genuine insight with errors and outputs too opaque to evaluate meaningfully. The Association for Human Mathematics responded by urging mathematicians to stop collaborating with OpenAI altogether, framing the issue as a threat to the discipline itself.
Now Alpoge—who worked closely with OpenAI tools during that earlier controversy—is claiming a result that, if true, would dwarf everything that preceded it. The pattern is impossible to ignore.
Timing is not coincidence
The announcement landed on November 20, the day after Anthropic’s historic initial public offering. Alpoge acknowledged the optics. “I realize the timing may attract cynical eyes,” he wrote. He also said he consulted Anthropic CEO Dario Amodei before posting.
That disclosure matters. It means the company was aware this claim was going public immediately around its most financially significant moment of the year. Whether that creates a conflict of interest is for the community to decide. What is clear is that the incentive structure around this announcement is unusual even by startup standards.
What happens next
The Riemann hypothesis has resisted proof since Bernhard Riemann posed it in 1859. It states that all non-trivial zeros of the Riemann zeta function lie on the critical line with real part one-half. Its resolution would have consequences across number theory, cryptography—public-key systems like RSA depend on properties of prime numbers—and quantum computing. The Clay Institute is holding seven Millennium Prize problems, each worth one million dollars, and the Riemann hypothesis is the most famous.
Recognition requires two things: publication in a refereed journal and at least two years of scrutiny by the global mathematics community. Neither clock has started. The manuscript has not undergone peer review. No independent team has published a reconstruction or critique. The Lean file exists, but formal verification of a proof of this breadth has never been attempted at this scale, and the community does not yet know whether automated checking will confirm every logical step.
Alpoge’s claim also extends to the generalized Riemann hypothesis, which would multiply the impact significantly. That is an enormous scope for a single submission, and enormous claims demand extraordinary evidence—not blog posts and website uploads.
The verdict is not in
None of this proves the claim wrong. None of it disproves anything. What it establishes is simple: no one responsible for evaluating such a claim is prepared to call it valid yet, and several prominent mathematicians think the burden of proof remains very high.
The safer reading of Alpoge’s own description is that Claude generated material that appears consistent with a proof strategy but traverses territory neither human nor machine has fully mapped. That is genuinely interesting. It also falls well short of “proof of the Riemann hypothesis.”
If the result eventually holds up—after peer review, after replication, after two years—mathematicians will celebrate. Right now, the prudent position is exactly what the available evidence supports: wait. Let the specialists do their work. And remember that in mathematics, the difference between a promising sketch and a proof is often a single unresolved lemma.