Independent mathematicians say a proof of the Riemann hypothesis, created using Anthropic’s Claude AI, is almost certainly correct. Researcher Andrew Sutherland confirmed the code works perfectly. One expert said the chance of anything being wrong is "approximately zero." Experts are now finishing final checks to ensure the logic holds up.

Independent mathematicians, including M.I.T.'s Andrew Sutherland, have compiled an AI-assisted proof of the Riemann hypothesis in Lean and say it is almost certainly correct.

Independent mathematicians now believe a proof of the Riemann hypothesis, produced with the help of Anthropic's Claude AI model, is almost certainly correct. One researcher who checked it said the chance of anything being wrong is "approximately zero." A final check is still pending.

The claimLevent Alpöge, a mathematician at Anthropic, said he had used Claude to solve the Riemann hypothesis. Many consider it the most important open problem in mathematics. The idea dates to 1859, when Bernhard Riemann suggested that a hidden pattern sits behind how prime numbers are spread out. Primes are the numbers divisible only by themselves and 1.

How the proof was checkedAlpöge first tested his own work by writing the proof out in Lean, a programming language. When a proof written in Lean compiles without errors, it is usually taken as a sign that every step of the logic and maths holds up.

That left some open questions, and Scientific American reported that researchers have since worked through them. The first was whether the proof really tackles the Riemann hypothesis, or a slightly different statement. Kevin Buzzard of Imperial College London explained the concern in a blog post. In his words, the first thing to ask about a PDF that comes with Lean code is: "what is the statement of the theorem in the Lean code, and does this correspond to the actual theorem statement."

The second question was whether other mathematicians could get the code to compile too. Independent researchers have now done that. They include Andrew Sutherland of M.I.T., who compiled Alpöge's result on his own without any problems.

What Sutherland foundSutherland said the Lean kernel had independently re-checked the proof using only the three standard axioms. "Honestly the chance that anything is fishy here is approximately zero," he said.

One check still to comeResearchers are not finished. They are still making sure the proof contains no hack, dodge or shortcut that quietly changes the meaning of the original hypothesis or of any other mathematical rule. This step is called a "comparator run," and it normally takes a matter of hours.

Buzzard wrote on an online chat server for the Lean community that a successful run is still awaited. He said it would confirm two things. One is that the code does not do anything nefarious, such as redefining the symbols in the statement. The other, if it is run in what he called "paranoid mode," is that the code compiles with several independent kernels, so it is not exploiting an unknown bug in Lean's official kernel.