Humans are known to detect probabilistic regularities in learning materials. In language, this ability is particularly well-evidenced for the auditory modality. In two experiments, we tested if learners rely on statistical regularities also when processing text. Italian speakers performed a lexical decision task on a large set of letter strings, distinguishing between nonword strings and strings that were words in a novel language. Participants successfully detected that novel words complied with the statistics of the Italian lexicon, and used this information to perform the task. Specifically, participants tracked the frequency of single bigrams, rather than estimating the average frequency of all bigrams in a string (Experiment 1). Interestingly, they did so even when the informativity of the latter variable was increased (Experiment 2). Our study constitutes the first large scale investigation of the relationship between statistical learning and visual word acquisition. We discuss how general visual statistical learning mechanisms may be used in reading.