Doesn't this just move the need to be smarter from the model to the harness - if a human sometimes can't tell whether a model has produced something correct or just mostly correct-looking BS, how can an automated harness do it?

OTOH, if the goal is simple ("break into a protected system") rather than more complex ("write an application that satisfies all requirements on all supported devices/screen resolutions etc."), that's of course more suitable for a harness.