Three experiments examined the role of causal expectations in sentence integration. Subjects read narratives that did or did not elicit strong expectations about upcoming events (high-expectation vs. low-expectation texts). Subjects were faster to judge a target sentence to be unrelated when it followed a high-expectation text than when it followed a low-expectation text. Subjects were slower to read an unimportant, expectation-violating sentence embedded in a high-expectation text. Subjects were faster both to read and to judge an expectation-fulfilling target sentence when it followed a high-expectation text. Three models of the generation and use of expectations are considered. The results support a model in which expectations are generated selectively. Once generated, expectations become the focus of upcoming sentence integration processes such that correct expectations help and incorrect expectations interfere. A third experiment found no evidence that expectations took the form of highly specific predictions.