A vast amount of literature suggests that the language processor generates expectations about upcoming material. Several studies have found evidence for a prediction error cost in cases where the comprehender encountered not the predicted word but a plausible unexpected continuation instead. This cost is argued to be a result of an inhibitory process that suppresses activation of the originally predicted word. Other studies, on the other hand, have found no such evidence for a prediction cost. In a probe recognition memory task, we find evidence for interference from an incorrectly predicted word, and in a self-paced reading study, we find evidence for facilitation when the originally predicted word is encountered later on in the sentence. Taken together, our results provide evidence against a strong version of the suppression account, in which all incorrectly predicted words are inhibited. Instead we argue in favor of a passive lingering activation account, in which activation for the disconfirmed prediction gradually decays over time.