Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs
A Lean proof certifies the formal statement and proof that the checker sees; it does not, by itself, establish that an AI faithfully translated the intended natural-language argument. This paper examines semantic mismatches in AI autoformalisation, including examples connected to OpenAI’s announced Navier–Stokes result.
The authors argue that ambiguity makes faithful translation deeply difficult, but the paper’s strong framing should be read carefully: a mismatch between a natural-language explanation and its Lean formalisation does not alone show that the formal proof is invalid or that the theorem is false. Lobsters commenters likewise cautioned against treating the result as evidence that LLMs cannot prove theorems; they debated how much human review a very large formal proof can realistically receive, and whether proof review should focus on the formal statement and key components.