Six experiments examined a bias in the probability and the time course of predictive inferences as a function of individual differences in emotional proneness and the adaptive importance of stimuli. Participants high or low in trait anxiety read context sentences predictive of potentially threatening events or nonthreat events. The contexts were followed by a target word to be named, which represented the predicted event or an unlikely event. Context constraints, the context presentation rate, and the interval between the context and the target word were manipulated. Evidence for inferences involved facilitation (shorter latencies) in naming the inferential target words following the predicting context, relative to when these words followed a control context. Predictive inferences were activated when there was congruence between anxiety and the threat content of stimuli, regardless of context constraints. The time course of predictive inferences was constant across context presentation rate; they occurred with considerable delay (1,050 ms) after the context but not earlier (i.e., 50 or 550 ms). We discuss the relevance and the integration of these findings with current models and research on inference processing.