Rational inference over a noisy channel can potentially explain readers’ occasional misreading. We tested if the prior probability of an intended message modulates the rate of misreading a transposed-word sentence as grammatical. In Experiment 1 we manipulated the cloze probability of a word given its full context (Because my socks had holes, I bought a new pair/pack) but found no reliable effect on the rate of noticing word transpositions (pair new vs. pack new). In Experiment 2 we manipulated the 4-gram frequency of the sequence ending with the transposed words and again found no effect (I always know what mean they vs. love they). We conclude readers do not effectively exploit full-context information to derive nonliteral messages. Despite the results of Experiment 2, comparison of error rates across conditions in several experiments suggests a role for local ngram statistics, though perhaps only in a restricted range of ngram frequency.