On 20 November 2026, mathematician Levent Alpöge released a 17-million-line Lean library claiming a proof of the Riemann hypothesis—one of the most enduring problems in mathematics, first proposed in 1859. The Lean and Mathlib communities have since confirmed that the proof compiles correctly, states the correct conjecture, and has been validated by multiple independent proof checkers, including Lean’s official kernel and third-party tools such as con-leche and con-ron. This verification marks a pivotal moment in formal mathematics, where correctness is no longer solely dependent on peer consensus but on machine-checked rigor.
Formal verification meets century-old conjecture
The Riemann hypothesis concerns the distribution of prime numbers and posits that all non-trivial zeros of the Riemann zeta function lie on the critical line with real part ½. Its resolution would have profound implications for number theory, cryptography, and computational complexity. While many claimed proofs have surfaced over the past 167 years, none have withstood scrutiny—until now, according to the Lean community. The proof’s formalisation in Mathlib, a comprehensive library of verified mathematics, ensures that every logical step adheres to axiomatic foundations. The statement RiemannHypothesis is already part of Mathlib and had been vetted by experts in number theory and formal verification prior to Alpöge’s contribution.
AI as collaborator, not author
Alpöge, employed by Anthropic at the time of the announcement, described in an essay how he used Claude—an advanced AI model—to explore mathematical ideas during a “stream-of-consciousness” process. He credited the AI with helping him navigate complex conceptual terrain, particularly near the end of the proof, where he said the model “flew, and danced, and sparkled.” However, the final proof was authored entirely by Alpöge and expressed in Lean, with no direct AI-generated code. The collaboration highlights a growing trend: AI as a cognitive partner in high-level mathematical discovery, not a replacement for human insight. This contrasts sharply with a separate, fully human effort by Stanford postdoc Alessandro Masullo, who—supported by professor Ravi Vakil—released a draft proof of the General Riemann hypothesis, explicitly stating it involved no AI.
Dual claims, divergent paths
While Alpöge’s proof has been machine-verified, Masullo’s remains unformalised and thus unverified. Vakil explained that the plan had been to spend a year refining the manuscript before public release, but the timing of Alpöge’s announcement prompted an early publication. Masullo called his work a message of resilience: “we are still here. We are very much alive.” The juxtaposition underscores a tension in modern mathematics—between formal, verifiable results enabled by computational tools and traditional, narrative-driven proofs rooted in human intuition. For now, only Alpöge’s proof meets the gold standard of formal verification.
What it means
For UK, US, UAE, and Kazakh tech and academic sectors, this event signals a shift in how foundational knowledge is validated. Organisations relying on cryptographic systems—especially those based on number theory—should monitor developments, as a proven Riemann hypothesis could influence future algorithmic design, though no immediate practical breakage is expected. The use of Lean reduces dependency on single experts and increases trust through transparency and reproducibility. For startups and research labs, investing in formal methods may offer long-term advantages in credibility and collaboration. The AI-assisted nature of the discovery also raises questions about intellectual ownership and reproducibility—though in this case, the proof stands independently of the tooling used to conceptualise it.
The mathematics community now faces the task of translating Alpöge’s formal proof into human-readable exposition. Until then, confidence rests on the machine verification, not intuitive understanding. As Andrew Sutherland and others have confirmed the compilation and logical consistency of the proof, the burden shifts to explanation—not validation.