Abstract Many languages contain apparently arbitrary word classes, of which the best known examples are the gender classes of languages such as French, Hebrew, and Italian. However, a number of phonological or semantic similarities typically exist among subsets of class members which could assist the language learner in discovering the class distinction. The primary objective of this study was to investigate the possibly facilitative role of phonological similarities among a subset of class members in the acquisition of otherwise arbitrary word classes, using artificial miniature languages. The learning of an experimental language in which a subset of the members of each noun class shared a set of phonological features (i.e., a common ending) was compared with the learning of a control language in which phonological similarities were randomly distributed between the two noun classes. Many adult subjects learning the experimental language acquired the word classes while no one learning the control language did. The results were replicated with nine- and ten-year-old children. These findings support an acquisition model in which inherent similarities among words are the means by which the classes are discovered; they cast doubt on models with mechanisms for acquiring completely arbitrary classes.