AI Claims Riemann Hypothesis Proof: The Three Gates It Must Pass
Anthropic mathematician Levent Alpöge claims an AI-assisted Lean proof of the Riemann Hypothesis, but the 250,000-page formalization must clear three hurdles—compilation without gaps, correct theorem statement, and human comprehensibility—before mathematicians accept it as a valid proof.
What the Riemann Hypothesis Says
The Riemann Hypothesis governs the distribution of prime numbers. While primes appear irregular, their count up to a given limit is well approximated by the logarithmic integral. For example, up to 1 million there are 78,498 primes; the logarithmic integral estimates 78,627.5, an error of about 130. If the hypothesis holds, this error stays within a narrow bound: Schoenfeld (1976) showed the error at 1 million cannot exceed roughly 550. The hypothesis is equivalent to stating that all non-trivial zeros of the Riemann zeta function have real part 1/2 — i.e., they line up on a single vertical line. In Lean's Mathlib library, the hypothesis is expressed in a single line of code.
This definition is crucial for the second verification gate.
Lean: A Rule‑Checking Proof Assistant
Lean is both a programming language and a proof checker. You write a proof as code; Lean verifies each step follows from previous conclusions using allowed rules. Missing or jumped steps are rejected. In the official online editor (live.lean-lang.org, Lean v4.37.0‑rc1 + Mathlib), 2 + 2 = 4 compiles, while 2 + 2 = 5 produces a type‑mismatch error. Mathematicians trust Lean because its kernel is tiny and can be independently audited; a 250,000‑page derivation can be checked exhaustively by machine.
Gate 1: Compilation ≠ No Cheating
Lean provides sorry to skip a proof step. Code with sorry still compiles (with a yellow warning). The author tested a six‑line file that proves the false statement 1 = 2 using a single sorry — it compiles. To detect hidden sorry s, one must print the axioms the final theorem depends on. Under Lean's standard, only three axioms should appear: propext, Classical.choice, Quot.sound. Any extra axiom (e.g., sorryAx or a custom axiom) invalidates the proof. Alpöge states "no extra assumptions were introduced," but as of November 21, no one outside the author's circle has publicly confirmed a clean compilation of the 14 zip archives.
Gate 2: Proving the Right Theorem
The checker only verifies that the stated theorem is proved, not that the statement matches the intended problem. The author illustrates: define an even number as "divisible by 2 with remainder 0" and prove the sum of two evens is even — correct. Then accidentally change the divisor to 1; the proof still compiles and even proves "every number is even" because any number mod 1 is 0. This is a different error from sorry: the proof is complete, but the definition is wrong. For the Riemann Hypothesis, every definition (zeta function, zero, real part) must trace back to accepted mathematics. The simplest case is if the final theorem directly proves Mathlib's existing RiemannHypothesis statement; otherwise, someone must manually compare the custom definitions line by line. Lean's online editor has a "Can this proof be trusted?" button for exactly this cross‑check.
Gate 3: Machine Trust ≠ Human Understanding
Even if the first two gates pass, mathematicians will not celebrate immediately. The Four Color Theorem (1976) was accepted only after Gonthier's 2005 Coq formalization. The Kepler Conjecture (Hales, 1998) required a decade‑long formalization effort (completed 2014) because reviewers could only claim 99% confidence. In both cases, the machine settled "is it correct?"; mathematicians still needed "why does it work?" — new ideas that can be reused. A 250,000‑page derivation that cannot be compressed into a human‑readable paper remains a sealed black box. Terence Tao has publicly warned about such AI‑generated black boxes. Alpöge himself admits the proof timeline is rough, may contain suboptimal choices, and promises a human‑readable version later.
A Recent Precedent: The August 2026 Result
In August 2026, Anthropic announced an improvement of the lower bound for the proportion of zeros on the critical line from 41.6% to 67.2%. That release followed the proper template: a human‑readable paper, Lean formalization that passed the official comparison tool, and review by number theorists inside and outside the company. The announcement explicitly stated that Claude did not prove the Riemann Hypothesis and that the technique was not seen as a path to a full proof. Three months later, the full hypothesis claim appeared with code first and paper still pending.
When Would the Judgment Change?
The author's "not yet proven" stance has a clear falsification condition: if within weeks an independent party compiles the code cleanly, the axiom report shows only the three standard axioms, and the final theorem is exactly Mathlib's RiemannHypothesis, then the first two gates are satisfied. The third gate — human understanding — may still take years.
Simplifications Made
The three gates are presented sequentially for clarity; in practice they overlap (compilation and definition checking happen together, proof‑sketch readers spot code issues). The author has not personally audited the code; all facts come from Alpöge's note and public reports, and new developments are likely.
Trust‑Chain Model
A machine proof is trustworthy iff: (1) the checker kernel is reliable, (2) no extra assumptions are smuggled in, (3) the proved statement matches the original problem, (4) humans understand the argument. The overall confidence is the product of each link's confidence. As of November 21, links 2‑4 lack public verification. Illustrative calculation: 99% × 90% × 80% × 50% ≈ 36%. The weakest link dominates, so effort should focus there. Failure modes: a kernel bug or an error in the accepted mathematical definitions would require re‑evaluation of the entire chain.
Takeaway: Three Questions for the Next "AI Proves Century‑Old Problem" Claim
When you see such a headline, ask: (1) Has someone outside the author's circle independently compiled the code? (2) Which axioms does the final theorem depend on? (3) Does the formal statement match the original problem exactly? Until all three have answers, the claim is only a serious declaration worth watching. These questions apply only when formalized code is provided; if only a paper appears, fall back to traditional peer review and expert opinion.
Thus, the Riemann Hypothesis is not yet proven. Machines are turning "is it correct?" into a checkable task; but "is it the right theorem?" and "do we understand it?" still require human judgment.
References
[1] Unite.AI. Levent Alpöge's Unverified Riemann Claim Puts Claude's Lean Code in Focus, 2026‑11‑20.
[2] OfficeChai. Claimed Riemann Hypothesis Proof Made With Claude Appears On Anthropic Mathematician Levent Alpöge's Website, 2026‑11‑20.
[3] Schoenfeld L. Sharper bounds for the Chebyshev functions θ(x) and ψ(x). II. Mathematics of Computation, 1976. The figures 78,498, 78,627.5, and error bound ~550 are computed from this bound.
[4] Gonthier G. Formal Proof—The Four‑Color Theorem. Notices of the AMS, 2008.
[5] Hales T, et al. A Formal Proof of the Kepler Conjecture. Forum of Mathematics, Pi, 2017.
[6] Anthropic. Riemann zeta (research announcement), 2026‑08‑10.
[7] Lean examples in this article were tested by the author in the official Lean online editor (live.lean-lang.org, Lean v4.37.0‑rc1 + Mathlib).
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
