Recent experiments (Beckers, De Houwer, Pineño, & Miller, 2005;Beckers, Miller, De Houwer, & Urushihara, 2006) have shown that pretraining with unrelated cues can dramatically influence the performance of humans in a causal learning paradigm and rats in a standard Pavlovian conditioning paradigm. Such pretraining can make classic phenomena (e.g. forward and backward blocking) disappear entirely. We explain these phenomena by a new Bayesian theory of sequential causal learning. Our theory assumes that humans and rats have available two alternative generative models for causal learning with continuous outcome variables. Using model-selection methods, the theory predicts how the form of the pretraining determines which model is selected. Detailed computer simulations are in good agreement with experimental findings.