Four tachistoscopic forced-choice recognition experiments explored the flexibility of processes underlying word perception. Stimuli were words, orthographically regular but unfamiliar pseudowords, and orthographically irregular nonsense strings. In the first two experiments, subjects knew that several different kinds of stimuli would occur in each block of trials and that one kind would occur much more often than the others. No matter which stimulus subjects expected to see most often, accuracy on words and pseudowords differed little, and both were identified considerably better than nonsense. In the third and fourth experiments, subjects were led to believe that only on stimulus type would occur but were surreptitiously shown another type on a small number of trials. Words were again identified more accurately than nonsense, and the size of the effect was independent of expectations. However, when either words or nonsense strings were expected exclusively, pseudoword accuracy did not differ from nonsense accuracy. Only when subjects knew that pseudowords would occur did they identify pseudowords more accurately than nonsense. This dissociation between word and pseudoword identification indicates the operation of two independent encoding mechanisms during tachistoscopic recognition, a stimulus-specific or logogenlike system sensitive to particular familiar strings and an orthographic mechanism sensitive to generally applicable constraints on letter sequencing. The stimulus-specific mechanism appears to be utilized automatically, but use of the orthographic mechanism is under strategic control. As shown in the first two experiments, however, rather extraordinary measures were required to demonstrate the flexibility of the orthographic processes used in this task.