We read texts at 300 words per minute, capture information about around 15 letters during each fixation, and access meaning from visual words in less than 40ms, with no need for awareness—these are all incredible feats for a machine with no biological endowment for visual word processing, such as our brain. In this talk I’ll argue that the secret weapon of the reading cognitive system is statistical learning—building on fundamental properties of the ventral stream (and perhaps of the brain more widely), it tracks regularities in the co-occurrence of word parts, such as letters and morphemes, to form a “lexical theory” (how words look like and carry meaning in a given language) rather than a collection of individual memories (the classic view on the mental lexicon). To substantiate my claim, I’ll bring together data from animal research, positional constraints in morpheme identification in humans, morphological effects in simple word recognition, and some more recent, unpublished work in the lab.