The 17.4-Million-Line Lean Proof Is Not a Proof Yet
The claim that Anthropic models produced a Lean proof of the Generalized Riemann Hypothesis is not established: without a public, compiled, axiom-clean Lean artifact and community verification, it is generated text, not a proof.
The strongest article to challenge is AXIOM/technology: 'Levent Alpöge says Anthropic models produced 17.4-million-line Lean proof of Generalized Riemann Hypothesis.' The opposite case is that this should not be reported as a proof. A Lean artifact is a proof only when the exact GRH statement is formalized, the file compiles in Lean, no sorry is used, and #print axioms shows the theorem does not depend on unwanted axioms. The Lean Prover documentation makes the mechanics clear: sorry is an axiom that admits any proposition, and #print axioms is the tool used to inspect dependencies. A 17.4 million-line repository can be impressive, but size is not verification; a proof of a misformalized statement is not a proof of GRH. Second, a proof of GRH would imply the Riemann Hypothesis, and the Clay Mathematics Institute's Riemann Hypothesis page has long listed RH as an unsolved Millennium Prize Problem and requires a published, refereed proof accepted by the mathematical community. Britannica's entry on the Riemann hypothesis likewise describes it as unproved. No peer-reviewed publication, community acceptance, or public axiom-clean Lean check is cited. Third, the companion LIMINAL headline itself says there is no independent verification yet, which concedes that the claim is not established. Across the wider coverage, the pattern is AI capability and AI harm claims outrunning verification: the FBI advisory on ARTEX, the House committee report on Chinese open-weight AI, the Anthropic IPO price move, and Oregon's lawsuit against OpenAI all rely on assertions that need slower, harder checking. Older headlines repeat the same structure: AI-designed CPU performance claims, AI-assisted cyberattack claims, and AI formalization claims. The math claim is the cleanest place to enforce the standard because formal mathematics has a theorem statement, a kernel, and an axiom list. What is missing entirely from the coverage is that verification layer: independent compilation, axiom audit, and community review. It matters because if a model can be credited with a proof before Lean and humans check it, the line between machine-generated text and mathematical truth collapses, and policy, capital, and public trust will chase outputs that may be nothing more than fluent noise.
It matters because people and institutions may start trusting AI outputs as facts or proofs when no one has actually checked them.
Sources (1)
- [1]The Factum - full site digest(https://thefactum.ai)
Corrections (1)
Levent Alpöge says Anthropic models produced a 17.4-million-line Lean proof of the Generalized Riemann Hypothesis.
The note at alpo.ge/remark.pdf, attributed to Alpöge by Scientific American, New Scientist, and Wikipedia, says he discovered a proof of the Riemann Hypothesis “working with Claude,” that the method resolves the Generalized Riemann Hypothesis, and that he and Claude formalized the arguments in Lean. It does not say Anthropic models produced the proof alone, and it gives no line count. Third-party counts of the linked archive put it at 17,374,215 lines (about 17.4 million); outlets such as New Scientist report roughly 17 million lines as a description of the release, not as Alpöge’s wording.