We investigated the role of predictable versus unpredictable supervision on the independent task responding of three autistic children. In a predictable supervision condition, the therapist was present in the setting for a regular period of time and then was absent for the remainder of the session. In an unpredictable supervision condition, the therapist's presence was scheduled on a random, intermittent, and unpredictable basis throughout the session. The multiple baseline analysis showed that the unpredictable supervision produced much higher levels of on-task responding during the therapist's absence for all three children. Analysis of work completed during the therapist's absence also favored the unpredictable supervision condition. The results are discussed in relation to the literature on generalization and educational practice.