Emerald Icon
◆

Emerald Pages

◆
placeholder

Photo: Emerald Book Image

For 167 years, the Riemann hypothesis has stood as the Mount Everest of mathematics — a deceptively simple conjecture about prime numbers that has resisted every human attempt to conquer it. So when Anthropic mathematician Levent Alpöge announced on November 20, 2026, that he and the company's Claude AI had finally scaled the summit, the world took notice.

There's just one problem: almost no one can verify whether they actually made it. Alpöge's "proof" isn't a traditional mathematical paper. It's roughly 17 million lines of Lean code — a programming language used to mechanically verify logical arguments — backed by an underlying AI-generated text spanning more than 250,000 pages. That's approximately 500 full-length novels worth of mathematical argument, a volume so vast that Alpöge himself refused to publish it, calling it a "massive unreadable travesty of 'slop.'"

The code compiles. The computer says the logic is sound. But in mathematics, that's not the same as solved. Until independent mathematicians audit the setup and agree there are no loopholes — a process that will likely take years — the Riemann hypothesis remains officially unsolved, and the Clay Mathematics Institute still lists it as open.

The Verification Gap

Here's the reality at the heart of the Anthropic announcement: computer verification and mathematical truth are not the same thing. When OpenAI announced its own AI-generated solution to the Navier-Stokes Millennium Prize problem in September 2026, its Lean code compiled cleanly. Weeks later, mathematicians from King's College London and the University of Cambridge discovered that OpenAI's AI had subtly mistranslated its human-readable paper into the verification code. The computer had checked a slightly altered version of the theorem — not the one OpenAI actually claimed to prove.

The same trap awaits Alpöge's Riemann proof. While the 17 million lines of Lean code compile without errors, mathematicians still have to manually audit the definitions to ensure the AI didn't subtly alter the problem statement, smuggle in a hidden assumption, or exploit a loophole in the Lean system. The proof might be logically valid according to a computer long before any human actually understands how it works — or whether it proves the right thing at all.

A Proof No Human Can Read

The sheer scale of the proof creates a bottleneck that no amount of computing power can solve. Traditional peer review — the process by which mathematicians read, check, and challenge each other's work — is impossible when the document is 250,000 pages long. Instead, verification has to shift from checking the argument itself to checking the code that formalizes it. Mathematicians must verify a much smaller set of questions: Does the top-level theorem written in Lean accurately translate the actual Riemann hypothesis? Does the code rely only on standard mathematical axioms, without using any "cheat" commands like sorry in Lean that tell the computer to skip proving a step? And even if the computer confirms the logic is airtight, can anyone figure out why it works?

  • Statement Accuracy: Does the Lean code accurately translate the actual Riemann hypothesis, without subtly changing a definition or adding a hidden assumption?
  • Axiom Integrity: Does the code rely only on standard mathematical axioms, without using "cheat" commands that skip proving a step?
  • Conceptual Extraction: Even if the computer confirms the logic is 100% airtight, mathematicians still have to figure out why it works so they can compress 250,000 pages of brute-force AI output into human understanding.

This is the "verification debt" problem in action. AI labs can generate hundreds of thousands of pages of dense, brute-force math in weeks. Yet auditing just the top-level definitions, ruling out subtle formalization bugs, and extracting the core human ideas from 17 million lines of code still takes months or years of grueling effort from a very small pool of elite specialists. Instead of the author doing the work to make a proof clear and convincing for peers, massive AI dumps shift the entire burden onto the rest of the mathematical community to play detective and hunt for hidden flaws.

The Precedent of Failure

This isn't hypothetical. The OpenAI Navier-Stokes debacle showed exactly how AI-generated proofs can fail even when the computer says they're correct. Beyond the translation errors, fluid dynamics experts criticized OpenAI's proof for exploiting a loophole in the problem's wording — proving the equations "blow up" by introducing an artificial, finely tuned external force to stir the fluid, rather than proving it happens naturally without outside interference. The computer verified a version of the problem that wasn't actually the Millennium Prize question.

Alpöge and NYU mathematician Tristan Buckmaster rushed out their own AI-aided Navier-Stokes proof on the same day as the Riemann announcement, claiming to close OpenAI's loophole by proving the equations blow up naturally with no external force. But because they rushed to publish out of fear of getting scooped again, their 79-page proof has not yet undergone independent peer review or full formal verification. The pattern is clear: AI can generate plausible-looking mathematics at software speed, while human verification remains stuck at traditional academic speed.

What makes the Anthropic announcement so easy to dismiss is that it follows a well-worn pattern. The Riemann hypothesis attracts new "solutions" every single year — dozens of them, according to major mathematics journals, preprint servers like arXiv, and the Clay Mathematics Institute. Most are quietly rejected by journal editors without fanfare. Only a handful ever make headlines, and so far, every one of them has failed.

In 1945, the respected German-American mathematician Hans Rademacher announced a proof that initially fooled even the legendary number theorist Carl Ludwig Siegel. Just as The New York Times and academic journals were preparing to celebrate, Siegel found a subtle flaw in Rademacher's reasoning. In 2004, Louis de Branges — famous for proving the Bieberbach conjecture two decades earlier — posted a 124-page proposed proof online. Mathematicians quickly identified critical gaps, though de Branges kept revising for years. In 2008, Xian-Jin Li's arXiv proof generated brief excitement before peer scrutiny uncovered a fatal error within days. And in 2018, at age 89, Fields Medal and Abel Prize laureate Sir Michael Atiyah delivered a lecture claiming a "simple" five-page proof; mathematicians concluded it was vague and lacked rigorous steps.

Princeton mathematician Peter Sarnak once noted that roughly 90% of the proposed proofs submitted to top journals don't even use the actual properties of prime numbers — meaning they rely on generic mathematical tricks that cannot possibly capture the deep arithmetic structure of the problem. Because the Riemann hypothesis looks deceptively simple on the surface, it acts like a magnet for false alarms.

  • Hans Rademacher (1945): Fooled even Carl Ludwig Siegel at first — until Siegel found the flaw.
  • Louis de Branges (2004): A 124-page proof with critical gaps, revised for years afterward.
  • Xian-Jin Li (2008): Retracted within days after a fatal error surfaced.
  • Sir Michael Atiyah (2018): A five-page "simple" proof that mathematicians found vague and incomplete.
  • Levent Alpöge & Claude (2026): An AI-assisted proof via 17 million lines of Lean code — currently unverified.
  • Alessandro Masullo (2026): A competing 940-page all-human proof announced hours later, still undergoing major revisions.

So when Anthropic announced its quarter-million-page AI proof, mathematicians had every reason to be skeptical. The claim arrived with more computational firepower than any previous attempt — but the same unresolved question: has anyone actually checked whether the proof proves the real thing?

What Would It Take to Actually Solve It?

For the Clay Mathematics Institute to officially declare the Riemann hypothesis solved and award the $1 million prize, the proof must be published in a recognized venue, survive two years of public scrutiny without anyone finding a flaw in the definitions, the compiler, or the logic, and gain broad consensus from the mathematical community that the problem has genuinely been resolved. Right now, none of those conditions have been met.

Mathematicians will use a "trust the kernel, audit the edges" approach — relying on the computer to check the 17 million lines of logic while humans manually inspect the problem's setup and try to reverse-engineer the core ideas. They'll scan for cheat commands, run the code through independent Lean type-checkers, and use AI tools to map out the 17 million lines like a giant dependency graph, isolating the key mathematical breakthroughs from the millions of lines of routine algebraic grinding. Only then can human mathematicians read and translate those core lemmas into a normal, human-sized academic paper.

But until that happens — and it will take years — every major AI proof claim is going to carry a massive asterisk. The Riemann hypothesis remains unsolved. Anthropic's proof remains an unverified claim. And the mathematical community is left wondering whether this is just the messy, bloated first draft of a breakthrough that humans can eventually distill down to 20 clean pages — or whether the hardest problems in the universe simply require proofs too giant for the human brain to hold.

For now, the answer to the question "Did Anthropic solve the Riemann hypothesis?" is the same as it was on November 19, 2026: No. Not yet. And possibly not ever, in any form a human can understand.

No Ads. By Us. For Us.

This article was made possible by readers like you. We hope it inspired you to support Emerald Book, so we can continue producing content like this.

We will never show you ads, sell your data, or require a subscription to consume our content. Your gift helps us keep the truth accessible.

Click the Support button to give a gift of any amount today.

Thank you for making this work possible.

◆

Emerald Pages is a publication of
Emerald Book, Inc.

Follow us
Share
Scroll to Top