Hoeks et al. (submitted) found that an unexpected word following a highly predictive sentence context elicited a significant increase in N400 amplitude, relative to when the target word concluded weakly predictive sentence fragments containing the same content words. This suggests that readers actively predict upcoming words and that this has the effect of making the processing of unexpected words more difficult. In the present experiment we investigated whether this finding can be replicated when participants read the sentence for comprehension, without having to give a plausibility judgment (as in Hoeks et al.). Additionally, we investigated whether the effect of prediction diminishes with increasing sentence length, manipulated by inserting a three-word prepositional phrase before the final word of the short sentences. The results showed no main effect of prediction when participants were asked to read for comprehension. The effect seems to be absent because the non-predictive condition is more negative in terms of N400 amplitude than in the earlier experiment, rather than because the predictive condition was more positive. Nevertheless, planned comparisons showed a small but significant effect of prediction in the longer sentences. Apparently if there is no expectation that sentences may be anomalous, participants may try harder to make sense of an implausible sentence, thereby possibly flooding the effect of prediction. We conclude that readers do use on-line prediction, but that the apparent costs of a mismatch may seem bigger than they actually are when readers have to make plausibility judgments.