Increasingly many findings suggest that the N400, long thought to reflect semantic processing, can be completely ‘blind’ to blatant semantic anomalies. For example, most studies of ‘role-reversed’ sentences such as “The fox.SUBJ the poacher.OBJ hunted” in verb-final constructions show that the verb elicits no N400 effect relative to the canonical word order control, despite the obvious anomaly. This kind of ‘dumb’ N400 has been taken to reflect a semantic illusion, and it contrasts sharply with the extensive evidence for ‘smart’ N400 effects, where its sensitivity to semantic/pragmatic anomalies is not reducible to simple lexical associations. In this talk I will propose that the smart vs. dumb N400 contrast reflects how quickly the processor generates expectations about upcoming input. I will show how it is possible to turn the dumb N400 into the smart N400 and provide new evidence regarding the time course for expectation generation during online sentence comprehension. I will also discuss how this interpretation of the N400 can give a unified account of diverse ‘dumb N400’ effects elicited in other contexts.