Who needs explanation and when? Juggling explainable AI and user epistemic uncertainty