This paper is concerned with the process of human reading as a high-level perceptual task. Drawing on insights from artificial-intelligence research--specifically, research in natural language processing and continuous speech understanding--the paper attempts to present a fairly concrete picture of the kinds of hypothesis formation and inference processes that must take place during the reading process. It makes the case that many more alternative hypotheses are formed and evaluated than one might expect and that the majority of such processing happens below the level of conscious introspection. Two text passages of different levels of difficulty are considered in detail, and the applicability of such artifical-intelligence insights to the modeling of human reading are discussed. (Author)