Skip to content

Has AI Solved Mathematics’ Most Famous Unsolved Problem?

In Simple Terms

An AI model from Anthropic has produced a computer-checked proof that may solve a famous question about prime numbers. Independent mathematicians have confirmed that the proof passes important checks, but one final test is still underway to make sure the computer was asked to prove the right thing. For now, the claim is promising, but not settled.

A Centuries-Old Question About Prime Numbers

The story goes back to 1859, when mathematician Bernhard Riemann proposed a hypothesis about how prime numbers are distributed. Prime numbers are numbers that can be divided evenly only by themselves and by 1. Looked at one by one, they seem scattered along the number line. But Riemann’s hypothesis points to a deep underlying order—a pattern that is not easy to spot by examining just a handful of primes.

The problem matters for more than its appeal as a mathematical puzzle. It asks a broad question about the structure underlying prime numbers, which are fundamental to mathematics. If the hypothesis is proved, the result would be more than a technical accomplishment or a solution to a narrow equation. It would provide a rigorous argument about a pattern whose nature has been debated since the 19th century.

Still, the scale of the claim does not guarantee that it is correct. In mathematics, it is not enough for a result to seem plausible or for a long chain of reasoning to sound convincing. Every step must follow from what came before according to clearly defined rules, and the final conclusion must be the same one stated in the original problem. That distinction—between sound reasoning and correctly stating the question—has made the review of this work about more than simply reading the proof.

From Alpöge’s Claim to a Formal Proof

Mathematician Levent Alpöge, who is associated with Anthropic, announced that he had used the AI model Claude to produce a solution to the Riemann Hypothesis. The announcement drew widespread interest because the problem is one of mathematics’ best-known open questions, and because the work attributed to a language model was presented not as help with one calculation, but as a proof of the full result.

According to the information available, Alpöge did not stop at presenting reasoning in conventional mathematical language. He also wrote the proof in Lean, a computer language used to represent and check proofs under formal rules. In this approach, mathematical steps are expressed precisely enough for software to test them, rather than relying on a human reader to notice every detail or assume that a particular transition is valid.

This kind of verification is a powerful way to catch errors, such as a step that does not follow from its premises or a conclusion that the stated rules do not allow. If the code compiles in Lean, that means Lean’s engine has found no logical violation in the structure it was given. But that alone does not answer a crucial question: does the encoded structure really express the Riemann Hypothesis as mathematicians understand it, or does it express a different statement that looks similar but means something else?

Why Passing Lean’s Check Is Not Enough

The distinction is like the difference between checking the solution to a problem and first making sure the problem itself was copied correctly. Software can carry out its check with complete precision and confirm that a conclusion follows from the statement entered into it. But if that statement is not the hypothesis in question, passing the check does not prove that the original problem has been solved.

That is why mathematician Kevin Buzzard of Imperial College London focused on examining the theorem statement embedded in the Lean file and comparing it with the accepted formulation of the hypothesis. Evaluating a formal proof requires more than confirming that it compiles. Reviewers must make sure that the symbols and concepts in the code mean what they are supposed to mean, and that the statement being proved matches the announced mathematical goal. This is not a rejection of computer verification; it is a check of something that successful compilation alone cannot establish.

The issue is especially sensitive because software works with specific definitions and symbols, not with the intentions expressed by the person describing the proof in an announcement or explanation. If a symbol’s definition changes, or a statement is interpreted in an unexpected way, the system might appear to have proved the result while actually proving something else. That is why reviewers need to examine the meaning of the code, not just rely on a message saying that the program accepted it.

It helps to distinguish between two levels of confidence. The first is that the steps in the formal proof are consistent with the rules Lean applies. The second is that those steps, however consistent, address the right hypothesis in the intended form. Achieving the first level is significant, but it does not remove the need for the second. That is why the mathematical statement itself, and the way the code handles it, need an independent review.

Independent Checks Strengthen Confidence

Alpöge did not leave verification to his own run of the proof. Independent researchers checked the work, including Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology. They were able to compile the Lean proof independently without encountering problems at this stage. That matters because the result no longer depends only on the environment in which the code’s author wrote it or on his own experience running it.

Sutherland explained that the proof had passed Lean’s kernel check, and that the verification relied on only three standard axioms. In this context, axioms are accepted starting points within the mathematical framework, not results that the software added on its own. He said that, in his assessment, the chance of anything suspicious at this stage appeared to be extremely close to zero.

That assessment should be understood in light of what has actually been tested. Independent recompilation is strong evidence that the proof, as written in Lean, passes the kernel check under the relevant settings, and that the result does not depend on a single run by its author. But it does not automatically settle every question about how the theorem was formulated or how its symbols were defined. Researchers are still carrying out a closer comparison.

It is also important to note that this check does not replace human mathematical judgment altogether. The system can apply formal rules to the statement it receives, but specialists still need to confirm that the meaning of that statement matches the problem behind the claim. The verification therefore combines the machine’s ability to test many precise steps with mathematicians’ expertise in interpreting the theorem and deciding whether the proof really addresses it.

The Test That Is Still Pending

The next step those following the work are waiting for is known as a comparator run. The point of this procedure is to test the proof again, focusing on the possibility that the hypothesis as described and the statement implemented in the code do not mean the same thing. For example, it is meant to reveal whether the program changed the meaning of a symbol in the theorem statement or redefined it in a way that alters the claim being proved.

When run in a particularly cautious mode, the test is also intended to confirm that the code works with several independent kernels, not just Lean’s official kernel. Having multiple checking systems helps rule out the possibility that the proof passed because its code exploited an unknown flaw in the official engine, rather than because the proof itself was sound. Buzzard has said that passing one check is not enough if the goal is also to exclude this kind of possibility.

This does not mean researchers have found a flaw or trick in the proof. The information available does not report a discovered problem; it says that an additional test is still pending. There is an important difference between saying that the proof failed and saying that the verification process is not yet complete. So far, independent checks have produced positive results, but the technical questions raised by specialists have not all been resolved.

According to the available explanation, this kind of run usually takes a few hours. The fact that it usually takes only a few hours does not make it a formality, however: the test is designed to examine specific issues involving the integrity of the representation and the independence of the checking tools. Until it returns a successful result, the most accurate description is that the claim has passed important checks, while the comparator run is still needed to complete the picture.

What the Result Could Mean for Mathematics

If the proof passes the remaining checks, it would be a landmark example of AI being used to produce mathematical work that can be converted into a formal statement and tested in a reproducible way. One advantage of this approach is that evaluation does not depend solely on the reputation of the person announcing the result or on how persuasive their presentation sounds. Other researchers can run the code and check whether the proof passes an independent formal check.

But calling this an AI achievement should not obscure what mathematical work involves. Producing a proof, expressing it in Lean, and confirming that the formal statement matches the hypothesis are separate stages, and they should be treated separately. Even if Claude generated the reasoning, the result’s validity does not come from the model’s identity or writing ability. It depends on the proof passing rigorous checks and on the mathematical community agreeing that it establishes the claim it was meant to establish.

The broader importance of formal verification goes beyond this particular problem. A proof that others can rerun is easier to check reproducibly than one assessed through a single person’s reading. But this approach also brings an additional responsibility: reviewers must examine the definitions in the problem, understand exactly what the system checks, and remain alert to the limits of the tools involved. Confidence does not come from the word “passed” alone. It comes from knowing what passed, under which rules, and against which formulation.

Based on the information available, it is too soon to treat the Riemann Hypothesis as definitively settled. The story involves an extraordinary claim, supported so far by the independent checking of a Lean proof that relies on three standard axioms. At the same time, the comparator test has not been completed. Keeping that distinction clear is essential: major progress in verification should not become a final announcement before the checks requested by specialists are finished.

The Bottom Line

The proof of the Riemann Hypothesis attributed to Claude sits at the intersection of optimism and caution. Levent Alpöge has announced a solution, and independent mathematicians, including Andrew Sutherland of the Massachusetts Institute of Technology, have recompiled the Lean code without errors. Lean’s kernel has checked the proof using three standard axioms. These are strong signs that the formal proof is sound, but they do not by themselves settle whether its encoded statement matches the original hypothesis proposed by Riemann in 1859.

Specialists are therefore waiting for the comparator run, which examines symbol definitions and tests the code with independent kernels to rule out a change in meaning or reliance on a flaw in a single tool. Until that step is complete, the story is neither a fully settled announcement nor evidence that a flaw has been found. It is a significant claim supported by important checks and still undergoing its final review. An editorial note dated November 20, 2026, described the story as ongoing and subject to updates—a fitting description for a proof that is still awaiting completion of its verification.