Skip to content
The AI-First Web

Claimed Riemann proof with Claude raises the question of who checks AI work

An Anthropic mathematician says he has a proof. Nobody outside has verified it, and the prize rules add a two-year wait.

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Claimed Riemann proof with Claude raises the question of who checks AI work
In brief
  • Anthropic mathematician Levent Alpöge claims, with Claude's help, a proof of the Riemann hypothesis. The Verge reports it has not been independently verified.
  • The Verge describes mathematicians scrambling to digest a flood of AI-assisted results. The Clay Mathematics Institute's prize rules add a long public review period.
  • Leaders using AI for high-stakes work should ask who verifies its output, how long that takes, and what the organisation does in the meantime.

The claim is easy to make and slow to confirm

Anthropic mathematician Levent Alpöge says he has proved the Riemann hypothesis with help from Claude. The Verge reports that no one outside has independently verified the proof. It adds that Anthropic has made no comment and has not taken part in the announcement.

The hypothesis is among the seven Millennium Prize problems. Mathematicians first posed it in 1859. The Verge says it concerns how prime numbers are distributed, and how the zeroes of a certain function are involved. In plain terms, the problem asks where those zeroes sit, because that placement is tied to how primes are spread along the number line.

On his website, Alpöge writes that he believes he has discovered a proof with Claude. He also says the method extends to the Generalized Riemann Hypothesis.

Treat this as a claim, not a result. That distinction is the story.

The idea: who checks the work?

For centuries, the hard part of mathematics was finding the argument. If AI keeps lowering that cost, the question of who checks the result matters more. That is the publication's argument, not a finding of the sources.

The Verge offers one related observation. It describes mathematicians "scrambling to digest" a flood of AI-assisted findings. It does not measure how fast results arrive against how fast people can check them.

Companies will recognise the shape of the problem. A model can draft a contract, a code change or a financial analysis in seconds. A qualified person still has to confirm it is right.

How this proof is supposed to be checked

Alpöge says he and Claude "formalized the arguments without assumptions". The Verge reports the solution is presented in Lean. Lean is a proof-checking program: it accepts a proof only if every step follows from the one before. That is a machine-checkable form of evidence, and it differs from a persuasive write-up.

The sources do not say whether anyone outside has run or reviewed the Lean files. A formal proof also still has to be read by people. They must confirm that the statement it proves is the one mathematicians mean.

The Clay Mathematics Institute sets a second hurdle. Its rules, as The Verge describes them, require publication and then at least two years of surviving scrutiny before the prize can be considered. During that time the work must gain "general acceptance in the global mathematics community". The prize is $1 million per problem. Only the Poincaré conjecture has been officially solved since the prize was created 25 years ago.

That waiting period is a prize rule. It does not by itself show how hard checking is.

Two cases, two different questions

The Verge reports that OpenAI claimed in September to have cracked another Millennium problem, Navier-Stokes. It says the effort cost millions of dollars and used 10,000 agents. The result is widely considered credible but has not received formal recognition.

The Verge also reports allegations of questionable research practices. These include claims that OpenAI used unpublished work that mathematicians had shared with its AI tools. OpenAI has strongly denied wrongdoing. NYU professor Tristan Buckmaster claimed OpenAI used the route he and Alpöge had been pursuing. Alpöge had worked with Buckmaster on that problem.

The OpenAI case raises questions of credit and data provenance. The Alpöge case, in the sources, raises questions of verification and timing. The sources do not say the two share a cause, and they do not say speed caused the OpenAI dispute.

One reading, and it is ours, is that both cases show how much trust now depends on the process around an AI result. Two cases do not make a trend.

Other unresolved questions

Back in August, Anthropic described Claude work aimed at the Riemann hypothesis, yet cautioned: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." The Verge did not hear back from Anthropic on the new claim.

The Verge also reports that Alpöge texted CEO Dario Amodei about the solution "late Sunday night". It says his document suggests he held off publishing until after the company's IPO on Thursday. The sources do not say more than that.

What leaders should ask

You do not need to follow the mathematics to use the lesson. Where your teams rely on AI for work that carries real consequences, put these questions to them.

Who checks the output, and are they independent of the person who ran the model? How long does a proper check take compared with generating the work? Is there a machine-checkable test, as Lean is here, or does review depend on one person's judgement?

Also ask what the organisation does while a claim is unverified. A result announced is not a result confirmed. Our reading: the prize's multi-year review is a reminder that trust can take longer to build than an answer takes to produce.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Verge.