Social referencing consists of a child looking to the affective responses of an adult, which serve as discriminative stimuli for subsequent responding in the context of ambiguity or novelty. In this study, social referencing was defined as discriminative responding under a two-link chain. The discriminative stimulus for the first link was the presentation of experimental stimuli in the presence of which an observing response was required. Link 2 consisted of a conditional discrimination. The discriminative stimulus for the second link was an affective stimulus from one of two sets presented by the experimenter. Two experiments were conducted to teach children with autism to respond differentially to affective stimuli within the social referencing response chain, and to determine if differential responding generalized to similar stimuli. Experiment 1 attempted to evaluate discriminative responding to two sets of six affective stimuli in Link 2 of social referencing while participants encountered stimuli representing three types of tasks pictured in their activity schedules (i.e., handwriting, retrieving objects, and scripted social interaction). Because discriminative responding was not acquired by any of the three participants under that training paradigm, Experiment 2 was conducted. During this experiment, participants were seated at desks and were presented with stimuli that signaled social referencing. One affective stimulus from each of the two sets was used as the training stimulus. The remaining affective stimuli from the two sets were presented as probe stimuli to determine the extent to which each was part of an already established stimulus class. Participants were taught to engage in differential responding using manual guidance, differential reinforcement, and error correction. In the presence of an affective display from set 1 (e.g., smiling and nodding head), the correct response was a keep response in which the participants placed the stimuli in a bin on the desk. In the presence of an affective display from set 2 (e.g., shake head with eyebrows turned down), the correct response was a discard response in which the participants placed the stimuli in a garbage bin on the floor. Correct responding on training trials increased above baseline levels for all three participants with the systematic introduction of conditional discrimination training. Probe responding was inconsistent across the three participants, obviating analysis of stimulus class formation.