Abstract

How are biases encoded in our representations of social categories? Philosophical and empirical discussions of implicit bias overwhelmingly focus on salient or statistical associations between target features and representations of social categories. These are the sorts of associations probed by the Implicit Association Test and various priming tasks. In this paper, we argue that these discussions systematically overlook an alternative way in which biases are encoded, that is, in the dependency networks that are part of our representations of social categories. Dependency networks encode information about how features in a conceptual representation depend on each other. This information determines the degree of centrality of a feature for a conceptual representation. Importantly, centrally encoded biases systematically disassociate from those encoded in salient-statistical associations. Furthermore, the degree of centrality of a feature determines its cross-contextual stability: in general, the more central a feature is for a concept, the more likely it is to survive into a wide array of cognitive tasks involving that concept. Accordingly, implicit biases that are encoded in the central features of concepts are predicted to be more resilient across different tasks and contexts. As a result, the distinction between centrally encoded and salient-statistical biases has important theoretical and practical implications.