The present study explores how people learn about a causal system by interacting with it. Participants were given the task to identify the operation of virtual '"computer chips" by setting the value of various components and observing how those interventions influenced the setting of other components. Across conditions we manipulate the complexity of the causal system (i.e., number of nodes and connections), the number of alternative hypotheses (i.e. possible causal graphs) on each trial, and aspects of the "temporal stability" of the learning environment (if repeated interventions were made on a single, stationary system or if the system reset to different starting states following each intervention). Interventions were modeled by comparing them to an optimal Bayesian learner who chooses interventions to quickly reduce uncertainty about the structure. Our results suggest that naive Internet-recruited subjects choose highly informative interventions, but also deviate from the predictions of the optimal model in certain ways.