A connectionist model is described, which accounts for data showing a “mask appropriateness” effect in word and letter recognition. In a backward-masking paradigm, if the letter fragment masks that are used are of a similar size to the word or word fragment they are masking, performance is better than if the masks are longer or discontinuous (Jordan, 1990). The model simulates these effects by adding to the familiar architecture of the interactive activation model “length” units, which are stimulated whenever word units of a particular letter length are active, and “mask” units, which are stimulated by masks of particular sizes and in turn stimulate length units. Detailed predictions of the model are examined for both forward and backward masking, and are found to be in good accord with the data.