Flight is a testable characteristic (at least within certain bounds: a chicken can only fly a little, but it certainly flies more than a worm or a penguin).

What test would you propose where an LLM (or an RL agent containing an LLM) would reliably fail, but where an average human would reliably succeed?