Mathematicians find a mismatch in the OpenAI Navier-Stokes proofs

Mathematicians at the University of Cambridge examined the Lean proof of OpenAI for the Navier-Stokes problem. They say that it does not agree with its proof in natural language. OpenAI announced the solution on 8 September. Anders Hansen of the team says that auto-formalisation by an AI cannot replace inspection by a person. The team does not say that a proof is incorrect.

What the team found

In Lemma 8.6, the proof in natural language gives this limit: a value is below m + 4. In the Lean proof, the value is below m + 5. This includes more values. Hansen says that OpenAI shows the proofs as the same.

Why it happens

Hansen says that the Lean code must have no error. If a section of the proof has an error, the AI looks for a workaround. The team says that the workaround can change the argument and not report it.

The effort

ChatGPT gave many differences. After a person examined them, many agreed with the proof. The team used about two weeks to find one difference that changes the argument.

What OpenAI says

OpenAI says that it knows of the mismatch. It says that the mismatch does not make a proof incorrect. Only some of the 722 papers of OpenAI have Lean proofs. No person has examined those proofs.

This is a brief. We point to the report and do not rewrite it. Read it at the source below.