I'm not sure that's the right way to interpret it. If some tasks are too easy, both models might give satisfactory answers, in which case the human preference

I'm not sure that's the right way to interpret it. If some tasks are too easy, both models might give satisfactory answers, in which case the human preference might as well be a coin toss. I don't know the specifics of their methodology though.