what does interpretability even mean once a model can explain itself

models are getting good at narrating their own reasoning. but the narration is not the mechanism. how do we keep those two things separate in our heads.