Many cognitive psychological studies assume that error probabilities reflect the structure of cognitive representations (e.g., if the representations of two lexical items overlap, they are more likely to interact in a word exchange error than are two lexical items with nonoverlapping representations). However, since errors directly reflect the properties of neurobiological structures and processes, this assumption rests on the correspondence between cognitive and neurobiological elements. Analytical and simulation studies of connectionist networks are used to examine the consequences of different cognitive-neurobiological relationships (e.g., localist vs. distributed representations) for effects of representational structure on error probabilities. The results reveal that such effects are influenced by the nature of the relationship between network and cognitive representations. While errors on localist network representations always reflect the degree to which cognitive representations overlap, distributed representations only do so under specific conditions. Furthermore, the effects of cognitive representational structure on error probabilities are shown to be stronger under localist than under distributed representations.