ABSTRACT
A distributed connectionist network can account for both bookkeeping (Rothbart, 1981) and subtyping (Brewer, Dull, & Lui, 1981; Taylor, 1981) effects. The finding traditionally regarded as demonstrating subtyping is that exposure to moderate (compared with extreme) disconfirmers leads to subsequent ratings of the group that are less stereotypic. Despite learning that is incremental and analogous to bookkeeping, the simulations replicate this finding and suggest that the “subtyping” pattern of results will be drastically reduced if disconfirmers are encountered before the stereotype is well-established. This novel prediction holds with human participants and offers a tantalizing suggestion: Although moderate disconfirmers may produce more stereotype change, stereotype development might be discouraged by exposure to either extreme or moderate disconfirmers.
