If people do not understand when an AI system can fail, then accuracy alone is not enough. Trust has to be designed.