Intelligent messaging systems attempt to determine what information is important to a user’s task and when and how to interrupt the user with that information. Such systems are probabilistic, and will be wrong some percentage of the time. We conducted an experiment to assess the impact of variable reliability in a notification system on users’ trust and use of the system. We show how an unreliable system that violates users’ trust may lead to the abandonment of the system. This disuse behavior pattern persisted despite subsequent improvements in the reliability of the underlying intelligent system. Our results provide guidance in the design of notification user interfaces.