In a variety of domains, adults who are given input that is only partially consistent do not discard the inconsistent portion (regularize) but rather maintain the probability of consistent and inconsistent portions in their behavior (probability match). This research investigates the possibility that adults probability match, at least in part, because of two pragmatic assumptions they bring to the learning problem: (a) that the variation they see is predictable rather than random and (b) that their goal is to correctly learn that variation. Evidence from two experiments demonstrates that when either assumption is eliminated, people probability match less and therefore regularize more. These results are discussed with respect to age and domain differences in regularization.