As a last resort, do the code-review with a live pair programming session.

If they can't explain their own code then it is by default a bad pull request.

At the end of the day, everyone's time is being wasted on tokens and on the increasing cognitive complexity of AI generated code.