how does interpretability work