AI Riemann Hypothesis Proof: What Mathematicians Found

The headline says AI proved the Riemann hypothesis—but the actual result is narrower and far more interesting than that claim suggests. Most guides skip the crucial distinction, so this breakdown separates the proven zero-free bound from the famous unsolved conjecture and shows how experts tested the argument.

What the AI Riemann Hypothesis Proof Actually Claims

The reported AI Riemann hypothesis proof does not prove the Riemann hypothesis. Instead, it argues for a weaker result: nontrivial zeros of the zeta function stay away from the edges of the critical strip.

The Riemann hypothesis in plain English

The Riemann hypothesis says every nontrivial zero of \(\zeta(s)\) has real part

\[

\operatorname{Re}(s)=\frac12.

\]

In other words, the zeros should all sit on one vertical line in the complex plane. The conjecture remains unproved, even though more than 41% of zeros are known to lie on that line.[1][2]

What surprised me here is how easy it is to mistake “narrowing the possible region” for proving the final claim. Those are very different achievements.

What the Riemann zeta function and critical strip mean

The zeta function is deeply connected to prime distribution: its zeros help describe the fluctuations in where primes appear.[3] All nontrivial zeros are known to lie in the critical strip,

\[

0<\operatorname{Re}(s)<1,

\]

an open vertical band centered around the critical line \(\operatorname{Re}(s)=1/2\).[1]

Think of it like narrowing a search area on a map. You have ruled out large parts of the landscape, but you have not shown that every destination lies on one exact road.

Why the result is called the quasi-Riemann hypothesis

The reported AI argument claims a zero-free region near the strip’s boundaries. Its stronger version says nontrivial zeros lie within

\[

\frac18<\operatorname{Re}(s)<\frac78,

\]

creating the so-called “\(7/8\) wall.” A shorter proof reportedly establishes the weaker bounds

\[

\frac1{12}<\operatorname{Re}(s)<\frac{11}{12}.

\]

Because zeta’s functional equation gives symmetry around \(\operatorname{Re}(s)=1/2\), ruling out zeros on the right also yields a matching restriction on the left.[1][2]

But here’s the catch: neither bound forces zeros onto the critical line. The work may also extend to families of Dirichlet and certain Hecke \(L\)-functions, yet mathematicians still have to check every estimate, constant, boundary case, and hidden assumption before calling it a reliable advance.

Full RH vs. the Quasi-Riemann Bound

The cleanest way to understand the distinction is this: RH puts every nontrivial zero on one vertical line, \(\operatorname{Re}(s)=\tfrac12\). The quasi-Riemann result only says zeros cannot get too close to the two outer edges of the critical strip.

A visual comparison of the two claims

Think of the critical strip as the vertical band

\[

0<\operatorname{Re}(s)<1.

\]

RH compresses every nontrivial zero onto the center line:

\[

\operatorname{Re}(s)=\frac12.

\]

The claimed quasi-Riemann bound clears only the far-right section:

\[

\operatorname{Re}(s)>\frac78.

\]

By symmetry, it also clears the far-left section, leaving

\[

\frac18<\operatorname{Re}(s)<\frac78.

\]

That is a meaningful narrowing, but the central band remains broad. In my experience, this is where explanations often blur “a strong zero-free region” into “almost RH.” Those are not the same thing.

What the \(7/8\) wall means

The “\(7/8\) wall” means that no nontrivial zero crosses the fixed vertical boundary \(\operatorname{Re}(s)=\tfrac78\). The shorter argument reportedly gives the weaker boundary \(\tfrac{11}{12}\), or approximately \(0.9167\), while \(\tfrac78=0.875\) is the stronger result.[1][4]

But zeros could still exist anywhere between \(\tfrac12\) and \(\tfrac78\). So the wall is more like a fence moved inward from the edge—not a GPS route that has located every zero.

How the functional equation supplies the left-side bound

The functional equation pairs a zero at \(s\) with one reflected across the critical line:

\[

\operatorname{Re}(s)=\frac12.

\]

Therefore, ruling out zeros to the right of \(\tfrac78\) automatically rules them out to the left of \(\tfrac18\). The stated theorem applies not only to the zeta function, but also to Dirichlet \(L\)-functions and certain finite-order Hecke \(L\)-functions.[1][6][7]

That makes the result a notable advance in analytic number theory. Still, it is a zero-free-region theorem—not a solution to RH, which continues to demand \(\operatorname{Re}(s)=\tfrac12\) for every nontrivial zero.[2][3]

How Mathematicians Checked the AI-Generated Argument

A claimed proof of the quasi-Riemann hypothesis is not accepted just because its conclusion looks plausible—or because software can check pieces of it. Mathematicians have to audit every inference, lemma, dependency, and assumption, asking whether the argument really proves a zero-free region such as \(\operatorname{Re}(s)>\tfrac78\), rather than merely suggesting one.[1][2]

Logical dependencies and hidden assumptions

The review starts by tracing the proof from the definitions upward. Researchers check each Dirichlet series, Euler product, character sum, exponential sum, integral, and mean-value estimate, then verify that every later step depends only on results already established.

This matters because one unproved uniformity claim can quietly support an entire theorem. A formal system may confirm that lines of code follow logically, but independent mathematicians still need to understand whether the encoded statements match the intended analytic argument. As Peter Sarnak reportedly put it, a result should not be called a theorem until people have verified and understood it.[2]

Estimates, constants, and uniformity

Analytic number theory is especially sensitive to estimates that must hold uniformly across several parameters—height, conductor, character, field, or region in the complex plane. An estimate that works asymptotically, or only for “sufficiently large” values, may not be strong enough for the claimed conclusion.

Constants are the pressure points. Proving an error term is \(O(X^{0.9})\) does not automatically establish a boundary at \(7/8\) or \(11/12\); the exponent and its constant must fit every later inequality. What surprised me here was that the shorter companion proof drew attention precisely because it was easier to analyze, even though it established the weaker \(11/12\) bound.[3][4]

Boundary cases and exceptional parameters

Reviewers also test the argument where it is most fragile: close to the claimed \(7/8\) wall, near the critical line \(\operatorname{Re}(s)=\tfrac12\), and at small or exceptional parameter values. Functional-equation symmetry can transfer a right-side zero-free result to the corresponding left side, but only after the hypotheses are checked carefully.

So there are two standards: a proof can be logically valid, yet still be too opaque for independent audit. In my experience, clear exposition is not decoration in a 195-page analytic argument—it is more like a map: without it, even a correct route is difficult to verify.[5]

What Lean Formal Verification Does—and Does Not—Prove

How a proof assistant checks encoded mathematics

Lean is a proof assistant: it checks whether formally encoded steps follow from approved definitions, previously proved theorems, and its underlying logical foundations. Think of it less like a mathematician reading an argument and more like a very strict referee checking every move on a fixed chessboard.

The September version reportedly claimed that nontrivial zeros stay inside the \(1/8\)–\(7/8\) strip and included a Lean formalization.[1][2][3] In practical terms, Lean checked a formal artifact asserting results such as “if \(\operatorname{Re}(s)>7/8\), then \(\zeta(s)\ne0\).” The formal development reportedly involved roughly 486,000 lines across thousands of modules.[2][4]

The 199-page \(7/8\) proof

The manuscript itself ran to 199 pages, covering \(\zeta(s)\), Dirichlet \(L\)-functions, and some finite-order Hecke \(L\)-functions.[1][3] Its “\(7/8\) wall” is a zero-free boundary: no relevant zero appears to the right of \(\operatorname{Re}(s)=7/8\), with symmetry supplying the corresponding left-side bound.

What surprised me here was the contrast with the later, shorter version. The companion proof reportedly narrows the claim only to the \(1/12\)–\(11/12\) strip, spans 49 pages, and was not yet formally verified—the version mathematicians reportedly found easier to study.[1][2][3]

Why formal verification is not the same as understanding

Formal verification can catch an invalid deduction in the encoded proof. It cannot automatically explain why the strategy works, whether the exposition is readable, or whether the result is genuinely important. That distinction matters: machine-checked correctness belongs to the formal artifact, while human confidence still depends on interpreting its scope, assumptions, and mathematical significance.

There is also a basic trap. If the encoded target is misstated, Lean may verify the wrong theorem perfectly. So formalization is not a substitute for checking that the statement matches the intended quasi-Riemann hypothesis. This is where many tutorials get it wrong: a verified proof is reassuring, but understanding what was proved remains a human responsibility.

Why the Result Matters—and What Happens Next

A genuine improvement without a solved millennium problem

Moving a zero-free boundary from \(1\) toward the critical line at \(\operatorname{Re}(s)=\tfrac12\) is meaningful. A bound such as \(\operatorname{Re}(s)<\tfrac78\) limits how irregularly primes can be distributed, because zeta zeros influence the error term in prime-counting formulas. The reported shorter proof gives the weaker region \(\tfrac1{12}<\operatorname{Re}(s)<\tfrac{11}{12}\), while the longer claim reaches \(\tfrac18<\operatorname{Re}(s)<\tfrac78\).[1][3][5]

That is a real advance if correct—but it is not the Riemann hypothesis. RH says every nontrivial zero sits exactly on \(\operatorname{Re}(s)=\tfrac12\), not merely away from the edges of the critical strip. Think of it as narrowing the search area on a map, not pinpointing every address.

The role of existing analytic number theory

The reported arguments appear to recombine familiar machinery: zero-density estimates, large-sieve inequalities, bilinear forms, explicit formulas, and functional equations. The functional equation supplies the symmetry that turns a right-side zero-free result into the matching left-side bound.[3][4]

Earlier techniques associated with Alexander Dunn and Maksym Radziwiłł also reportedly influenced parts of the approach.[3] What surprised me here is that the interesting achievement may be organizational: an AI can assemble difficult tools into a coherent chain without inventing an entirely new theory.

What independent researchers still need to establish

The headline theorem still needs pressure-testing. Are every estimate and boundary case correct? Are the constants uniform across the claimed parameters? Does the proof cover exactly the stated family of Dirichlet and Hecke \(L\)-functions, rather than a narrower class?[3][4]

Researchers must also check dependencies between lemmas and determine whether the shorter proof can be formally verified. The longer \(7/8\) result reportedly has Lean formalization, while the shorter \(11/12\) argument was not yet formally verified in the cited reports.[3][5]

The accurate takeaway is measured: this may be a substantial quasi-RH result, but it does not put every nontrivial zero on the critical line. More broadly, it tests AI-assisted mathematics—and verification, clear exposition, and independent replication matter just as much as the theorem’s headline.