Although some types of inferences are mandatory for readers, predictive inferences (inferences for what will happen next) are generally considered elaborative or optional. Three experiments measuring probe word lexical decision latencies produced evidence for the online generation of predictive inferences during narrative text comprehension. Participants read passages that were designed to elicit inferences about what would happen next at a certain point in the story. Immediately following the point at which the inference was anticipated (the inference-generating sentence), a lexical decision probe word was presented that was related to the expected inference. Lexical decision latencies were shorter for probes related to the inference but only when the comprehension task encouraged prediction as a reading strategy. The results support the hypothesis that the production of predictive inferences is relatively strategy dependent.