Shanks (1991) reported experiments that show selective-learning effects in a categorization task, and presented simulations of his data using a connectionist network model implementing the Rescorla–Wagner (R–W) theory of animal conditioning. He concluded that his results (a) support the application of the R–W theory to account for human categorization, and (b) contradict a particular variant of contingency-based theories of categorization. We examine these conclusions. We show that the asymptotic weights produced by the R–W model actually predict systematic deviations from the observed human learning data. Shanks claimed that his simulations provided good qualitative fits to the observed data when the weights in the networks were allowed to reach their asymptotic values. However, analytic derivations of the asymptotic weights reveal that the final weights obtained in Shanks' Simulations 1 and 2 do not correspond to the actual asymptotic weights, apparently because the networks were not in fact run to asymptote. We show that a contingency-based theory that incorporates the notion of focal sets can provide a more adequate explanation of cue competition than does the R–W model.