Four experiments investigated how the reading of semantic associates of a predictive inference raised its activation level. Two kinds of semantic associations were distinguished: those that could causally relate text events to the inference in a representation of the described situation and those that could not. Lexical decision data indicated that in both cases the predictive inferences were activated to the same high level (Experiment 1), even when no other causal support of the inference could be found in the text (Experiment 3). However, the results of a judgment task performed at various delays (immediately after the reading of the predictive sentences, at a 3- sentence delay, and after the reading of all the texts) suggests that readers evaluate the plausibility of the predictive inferences on the basis of their causal support from text events. These findings suggest that the initial activation of predictive inferences mainly results from the associative constraints governing the elaborative process of the meaning of text words. The causal constraints of the predictive inferences intervene later to eventually reinforce the integration of the inference to readers' representation. The compatibility of these results with the comprehension models of Kintsch (1988, 1998) is discussed.