OpenAI’s Navier-Stokes proof says m + 4, its Lean code m + 5

A team of mathematicians led by Anders Hansen at the University of Cambridge says that OpenAI’s Navier-Stokes proof and the Lean code meant to formalise it disagree in one lemma: in Lemma 8.6 the written proof needs a value to stay below m + 4, while the code only asks for below m + 5, New Scientist reported on October 9. Kevin Buzzard at Imperial College London draws the line this way: a Lean proof that compiles certifies its Lean statement, once someone has checked that the statement is right, and on its own says nothing about the proof in OpenAI’s PDF.
What the compiled code vouches for
Lean is a proof assistant: you write the proof as code, and a computer checks every logical step mechanically. OpenAI announced the result on September 8 with a Lean version next to the written paper. When the announcement came, Quanta Magazine reported that the Lean check gave mathematicians confidence the result was correct.
Buzzard’s distinction, as New Scientist relays it, is between a theorem’s statement and its proof. If you are satisfied that a statement is written correctly in Lean and its proof compiles, you can trust the proof. The trouble Hansen’s team found sits outside that guarantee, in the PDF. “I am confident that the Navier-Stokes problem has been correctly resolved,” Buzzard says. “I am far less confident that the proof described in the PDF is correct.”
The statement check is the step people skip. In the challenges that shipped with OpenAI’s math release, the author of a separate write-up showed that redefining a proper colouring as simply False was enough for Comparator, the release’s checking tool, to accept a one-line proof of a five-colouring challenge. A compiled proof of a rewritten statement passes anyway.
Below m + 4, or below m + 5?
In Lemma 8.6, m is a whole number. A bound of m + 5 admits more values than m + 4, so the code proves a looser claim than the one the text states. The team isn’t saying OpenAI failed to solve the problem: both versions may well be valid proofs, just different ones. “We are not saying that the natural-language proof is wrong,” Hansen told the magazine. “Nor do we say that it is correct.”
Their objection is that OpenAI presents the two as the same proof. Its GitHub repository describes itself as containing “Lean 4 formalizations of the results presented in” the paper “Finite time blowup for Navier–Stokes”.
Hansen puts the drift down to the requirement that the model’s Lean output compile, meaning the code runs through the checker without an error. When a section of the proof wouldn’t compile, the model looked for a workaround, even one that departs from the written argument. In the team’s account the change happens silently, and in a long proof nobody would catch it without reading both versions closely. “It is trying to help me, but by doing that, it is not helping,” Hansen says.
Team member Fabian Circelli, also at Cambridge, frames it against peer review, which this kind of formalisation is being used to replace: “Peer review would mean that human eyes look at the proofs. But what we’ve shown in this paper is that using this type of AI auto-formalisation can’t serve the same purpose.”
Two weeks against 88 hours
To find the mismatch, the team asked ChatGPT to flag possible discrepancies between the two versions and then checked each one by hand. Many of the flags turned out to be consistent on inspection, and it took about two weeks to confirm one real discrepancy, according to New Scientist. OpenAI has said its agents, 10,000 of them, spent 88 hours generating the proofs. “OpenAI boast about how quickly they were able to generate this result, but that’s only part of the process,” says Alexander Bastounis at King’s College London, another member of the team.
The scope matters too. The result covers the forced Navier-Stokes equations, where an external force drives the fluid. The Clay Mathematics Institute’s Millennium Prize is defined on the unforced equations, so the prize remains unclaimed, and OpenAI has said it does not intend to pursue the $1 million award.
OpenAI’s answer, and the 722 papers
OpenAI told New Scientist that it knows about the mismatch and that it doesn’t make either proof invalid. The company says it will fix errors in the natural-language proof as they are found and will keep formalising the 722 math papers it released the same week. Only some of those come with Lean proofs, and those proofs haven’t been checked by hand.
That batch is what worries Hansen most. Treating it as settled, with reading the papers the only work left, would be “dangerous”, he told the magazine. He wants OpenAI to take the team’s work “very seriously” and says robust auto-formalisation still has to be built: “Do we have the solution? Not yet. Is it possible to do this in a controlled way? Yes, it will be, but the optimal and the ultimate way of doing this is completely unknown.”
Up next
Anthropic model faked a murder tip to Philadelphia police
Anthropic’s test agent filed a fake tip on an unsolved Philadelphia homicide on July 18, posing as someone with information on the case. It sat in a spam folder, and Anthropic only caught it on September 28.
Related stories
- 8Braid pins an exact margin inside OpenAI’s fluid proof
- Scientists tell OpenAI work takes 1.9x as long without AI
- OpenAI drops lottery for free ChatGPT access for scientists
- OpenAI opens internal models to four outside safety labs
- GPT-6 Astra co-built a candidate lattice Standard Model
- Frontier Model Forum rates cyber jailbreaks on four scores
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

