Manis et al. (1986, 1988) have suggested that, when a classification is superimposed on a series of items, this can lead to a reduction in the judged differences between the classes (interclass assimilation), whereas most previous research points to an accentuation of interclass differences. In a modified replication of conditions used by Manis et al., 82 subjects were presented with a series of vocabulary definitions, indicative of varying degrees of thought disturbance, supposedly provided by patients at different hospitals. For experimental but not control subjects, each definition was attributed to a patient from one of two hospitals, and the earlier items in the series were chosen so as to induce the expectation that patients from one hospital were more disturbed than those from the other. Subjects then compared pairs of midscale definitions (one from each hospital) and indicated which definition in the pair they considered more disturbed. Ratings of these test pairs by experimental subjects differed from those of controls in a direction of reduced discrimination between the classes, confirming the basic finding of Manis et al. This effect was not consistently influenced by the extremity of the induction series. Since the interclass assimilation effect found for test items also occurred for the induction items, our findings do not favour an interpretation in terms of midscale items being contrasted from other members of their class (within-class contrast). However, those subjects who were more confident in their ratings of the induction items showed less interclass assimilation. Implications for theories of category use and social stereotyping are discussed.