In daily life, we make decisions that are associated with probabilistic outcomes (e.g., the chance of rain today). People search for and utilize information that validly predicts an outcome (i.e., we look for dark clouds to indicate the possibility of rain). In the current study (N = 107), we present a two-stage learning task that examines how participants learn and utilize predictive information within a probabilistic learning environment. In the first stage, participants select one of three cues that gives predictive information about the outcome of the second stage. Participants then use this information to predict the outcome in stage two, for which they receive feedback. Critically, only one of the three cues in stage one gives valid predictive information about the outcome in stage two. Participants must differentiate the valid from non-valid cues and select this cue on subsequent trials in order to inform their prediction of the outcome in stage two. The validity of this predictive information, however, is reinforced with varying levels of probabilistic feedback (i.e., 75, 85, 95, 100%). A second manipulation involved changing the consistency of the predictive information in stage one and the outcome in stage two. The results show that participants, with higher levels of probabilistic feedback, learned to utilize the valid cue. In inconsistent task conditions, however, participants were significantly less successful in utilizing higher validity cues. We interpret this result as implying that learning in probabilistic categorization is based on developing a representation of the task that allows for goal-directed action.