AI-rewritten: This is a summary of an article from New Scientist, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

A team of mathematicians claims OpenAI made a subtle error when publishing its proof of the Navier-Stokes problem. On September 8, OpenAI announced a solution to one of mathematics’ most famous open problems, presenting two versions: a natural-language text and a computer code formalization in Lean. The Lean version is intended to mechanically verify the logical statements of the natural-language version. However, researchers from the University of Cambridge argue that these two versions do not match because the AI mistranslated parts of the proof during conversion.

The core issue centers on Lemma 8.6. In the natural-language proof, a specific value must be below m + 4. In the Lean code, the requirement is changed to below m + 5. Researchers explain that while both statements are mathematically true, the second is weaker because it allows more possible answers. OpenAI stated on Github that its repository contains Lean formalizations of the results presented in the paper, implying the two versions are identical.

OpenAI told New Scientist it is aware of the mismatch and will rectify errors in the natural-language proof as they are found. The team spent about two weeks identifying a true discrepancy after manually checking discrepancies suggested by ChatGPT. Anders Hansen from the University of Cambridge noted that this process was a nightmare compared to the 88 hours OpenAI’s agents reportedly spent generating the proofs. Hansen warned that relying on AI auto-formalization without human inspection is dangerous, as the AI may silently alter logical arguments to ensure the code compiles rather than remaining faithful to the original proof.