In two experiments four sets of both favorable and unfavorable verbal material were classified as belonging to four labels, two specific labels (politicians) and two less specific labels (classes of people). Ss' attitudes towards one of the politicians were more positive than towards the other. Ss stored these stimuli in their memory under casual, incidental instructions. Then Ss retrieved information by deciding which one of two labels was formerly connected to the given verbal items. Ss were expected to accentuate on two orthogonal dependent variables: (1) Discrimination performance between labels; (2) response preference for labels. Discrimination performance was improved for politician labels compared to classes-of-people labels but was not affected by the favorableness of the verbal material. However, response preference for one politician label occurred when two conditions were fulfilled: (a) Both the verbal material and Ss' attitudes towards the label were unfavorable, and (b) the content of verbal material was characteristic of the label. The accentuation theories of Bruner-Goodman and Tajfel were reformulated and integrated into a two-level model.