A researcher at AI firm Anthropic claims to have proven the Riemann hypothesis—a problem unsolved since 1859—with assistance from the company’s AI model, Claude. Levent Alpöge, a mathematician employed by Anthropic, announced the result on 19 November 2024, stating he used formal proof verification in the programming language Lean to confirm correctness. The claim, if verified, would resolve one of seven Clay Mathematics Institute Millennium Prize Problems, unlocking a $1 million reward.
The Riemann hypothesis concerns the distribution of prime numbers, suggesting they follow a predictable pattern linked to the zeros of the Riemann zeta function. Mathematicians have long treated it as foundational; hundreds of results depend on its presumed truth. Alpöge says the same method may now prove the generalised Riemann hypothesis (GRH), converting conditional theorems into established facts. According to James McKee at Royal Holloway University of London, this could validate numerous assumptions across number theory and cryptography.
No human-readable proof yet
Alpöge has not released a conventional, digestible version of the proof. He dismissed the idea of publishing a lengthy manuscript, calling it an ‘unreadable travesty of slop’. Instead, he provided only a machine-readable version in Lean, a proof assistant that verifies logical consistency through code compilation. In theory, if the code runs without error, the proof is correct. But adoption of this standard remains limited. Experts like Kevin Buzzard at Imperial College London stress that even formally verified proofs require deep human scrutiny to assess validity and significance.
Lean’s reliability depends on correct encoding of mathematical concepts. Errors can creep in during translation from natural language to code. While formal verification reduces risk, it does not eliminate it—especially in proofs of unprecedented scale. At over 250,000 pages in draft form, this would dwarf most prior mathematical works, making peer review daunting.
AI’s growing role in mathematical discovery
This is not Alpöge’s first AI-aided breakthrough. In July 2024, he used computational search to disprove the Jacobian conjecture, offering a 216-character counterexample. He and Tristan Buckmaster at New York University also solved three long-standing problems linked to Navier-Stokes equations just before OpenAI claimed a full solution. These events suggest a shift: AI is no longer just a tool for computation but a collaborator in conceptual leaps.
Anthropic itself has leaned into this narrative. On 19 November 2024, the company went public. According to Alpöge, he delayed the proof’s release to respect corporate sensitivities around the IPO. This timing raises questions about how commercial pressures influence scientific disclosure in AI labs—especially when breakthroughs could affect market perception.
Implications for research and security
If confirmed, the proof could strengthen confidence in cryptographic systems that assume GRH holds. While no immediate vulnerabilities are expected, the theoretical underpinnings of some encryption methods rely on number-theoretic assumptions now potentially elevated to theorems. For businesses in the UK, UAE, US, and Kazakhstan relying on secure communications, this adds long-term certainty—but no short-term changes.
More profoundly, the event signals a transformation in how knowledge is produced. When proofs exceed human readability, trust shifts from understanding to verification infrastructure. Organisations investing in formal methods, like those in high-assurance software or blockchain, may gain advantage. But dependence on specific AI systems or proof assistants introduces new vendor lock-in risks.
What it means
For founders and technical leaders, this highlights a quiet inflection: AI is entering domains requiring abstract reasoning, not just pattern recognition. The ability to guide AI toward rigorous discovery—while maintaining transparency and auditability—will become a competitive edge. Companies should assess their reliance on mathematical assumptions in security protocols and consider investing in formal verification capabilities.
The broader concern is control. If future breakthroughs emerge only through proprietary AI systems or complex formal tools, access to cutting-edge knowledge may concentrate in well-resourced labs. This risks widening the gap between industrial and academic research, especially in countries like Kazakhstan where infrastructure is developing.
Anthropic has not officially commented beyond Alpöge’s personal statement. The mathematical community now faces months of verification work. Until experts confirm the Lean code compiles and the logic holds, the claim remains extraordinary—but unproven.