This dissertation is concerned with the question of how people infer the consequences of active interventions in causal systems when only knowledge from passive observations is available. Causal Bayes nets theory (Spirtes, Glymour & Scheines, 1993; Pearl, 2000) provides a rational account of causality which explicitly distinguishes between merely observed states of variables (“seeing”) and identical states due to external interventions (“doing”), and which provides mechanisms for predicting the outcomes of hypothetical and counterfactual interventions from observational knowledge. By contrast, alternative models of causal cognition (e.g., associative theories) fail to capture the crucial difference between observations and interventions and thus are likely to generate erroneous predictions when the implications of observations and interventions differ. The basic research question of the eight experiments presented in this thesis was to investigate whether people who have observed individual trials presenting the states of a complex causal model can later predict the consequences of hypothetical and counterfactual interventions in a way predicted by causal Bayes nets. Consistent with the Bayes nets account learners were surprisingly good at inferring the consequences of interventions from observational knowledge in accordance with the structure and the parameters of the observed causal system. The experiments also show that participants were capable of taking into account the implications of confounding variables when reasoning about complex causal models. Although participants inferences were largely consistent with the predictions of causal Bayes nets, the studies also point to some boundary conditions of the competencies of lay reasoners. For example, learners had problems distinguishing hypothetical interventions from counterfactual interventions. In summary, the experiments strongly support causal Bayes nets as a model of causal reasoning. Alternative theories of causal cognition lack the representational power to express the crucial differences between observations and interventions and therefore fail to account for the results of the experiments.