Mathematician Uses Anthropic's Claude to Claim Solution to 167-Year-Old Prime Number Mystery

Mathematician Uses Anthropic's Claude to Claim Solution to 167-Year-Old Prime Number Mystery

2026-11-20 companies

San Francisco, Friday, 20 November 2026.
An Anthropic-affiliated researcher posted machine code claiming a proof of the Riemann hypothesis solved alongside Claude, sparking intense scrutiny over artificial intelligence transforming elite scientific discovery without traditional human peer review.

A Claim That Could Redefine Mathematics

On November 19, 2026, Levent Alpöge, a mathematician affiliated with AI lab Anthropic, published an unsigned four-page note titled Remark on his personal website (alpo.ge), asserting a proof of the Riemann Hypothesis and its generalized form using Anthropic’s Claude model as a primary collaborator [2]. The proof was released exclusively as Lean code, split into 14 compressed archive files, with no accompanying human-readable paper at the time of publication [2]. Alpöge’s announcement coincided with a separate claim, co-authored with NYU’s Tristan Buckmaster, regarding a proof that the Navier-Stokes equations can exhibit finite-time blowup without external forcing [2]. While the mathematical community has yet to validate the proof through peer review, the development underscores a pivotal moment in the integration of artificial intelligence into formal mathematical reasoning and discovery [2][4].

The Riemann Hypothesis: A 167-Year-Old Enigma

First posited by Bernhard Riemann in 1859, the Riemann Hypothesis is one of the seven Clay Millennium Prize Problems, each carrying a $1 million reward from the Clay Mathematics Institute [2][GPT]. The hypothesis concerns the distribution of the non-trivial zeros of the Riemann zeta function, which are deeply connected to the distribution of prime numbers [GPT]. Its resolution would not only resolve a centuries-old puzzle but also strengthen the foundations of modern cryptographic systems, which rely on the presumed intractability of prime factorization [GPT]. The stakes are immense: a proven solution would redefine number theory and have cascading implications for fields as diverse as quantum physics and cybersecurity [GPT].

AI as Co-Author: A New Paradigm in Mathematical Proof

The claimed proof’s reliance on Anthropic’s Claude model marks a significant departure from traditional mathematical practice. According to the anonymous author, the work was developed in collaboration with Claude over a 20-step timeline spanning from early October to November 16, 2026 [2]. The proof strategy reportedly employs a novel approach involving long chains of “theta lifts” instead of derivatives, drawing on advanced concepts such as relative trace formulas, Iwasawa theory, Bessel periods, and a weakened form of the Vandiver conjecture [2]. Rutgers mathematician Alex Kontorovich remarked that if a human had achieved this result, it would warrant an immediate Fields Medal, highlighting the perceived magnitude of the claim [2]. However, the lack of a human-readable document raises immediate concerns about transparency and verifiability, prompting widespread skepticism within the mathematical community [3].

A Community on Edge: Skepticism and the AI Proof Backlog

The announcement arrives amid growing unease about the rapid proliferation of AI-generated mathematical proofs. On October 6, 2026, OpenAI released a catalog of 722 AI-generated manuscripts, including a purported “quasi-Riemann hypothesis” and claims related to the Unique Games Conjecture and Hodge conjecture [5]. None of these results have undergone peer review, and OpenAI itself acknowledged potential issues in some unformalized proofs [5]. The Association for Human Mathematics criticized OpenAI’s release as a demonstration of power rather than scholarship, citing “total disregard” for scientific norms [5]. Meanwhile, arXiv introduced rate limits of two submissions per author per month in October 2026, following a surge in submissions that totaled 40,363 in September alone [6]. The backlog of unread papers now exceeds 722, with journals reporting submissions where authors cannot answer basic questions about their own work [6].

The Human Factor: Deskilling and the Future of Expertise

The rapid advancement of AI in mathematical discovery also raises existential questions about the future of human expertise. A 2025 study published in The Lancet Gastroenterology and Hepatology found that experienced endoscopists lost 6% of their unaided diagnostic skill within three months of routine AI assistance [7]. Similarly, a randomized controlled trial from the METR in July 2025 showed that senior software developers took 19% longer to complete tasks with AI tools, despite believing they were 24% faster [7]. These findings align with broader trends in professional deskilling, where reliance on AI erodes core competencies over time. Anthropic’s own employment report for October 2026 revealed a decline in junior technical hires from 27% to 9% over 18 months, suggesting a shift toward AI-augmented roles that prioritize oversight over foundational skill development [7]. The challenge now facing mathematicians—and professionals across disciplines—is whether the verification and contextualization of AI-generated results can be systematized before the pace of AI-driven discovery outstrips human capacity for scrutiny [7].

Sources


Artificial intelligence Cryptography