Lesioning studies are often used in cognitive neuroscience to make inferences about the architecture of cognition. Recently, computational models have been used to address some of the underlying assumptions-such as modularity and locality-often implicitly used when interpreting lesion data. In this article, we explore the ''functional localization'' assumption and its role in interpreting lesioning data, especially from double dissociations. The functional localization assumption states that units or subunits within an information processing system become functionally specialized for dealing with specific aspects of the input environment. Networks were trained on one of two problems-an abstract ''rules and sub-rules'' problem, and a more concrete ''logic classification'' problem-and then systematically lesioned. Networks were analyzed in terms of their overt behavior, and more importantly, in terms of their internal structure. Performance deficits in both form and magnitude could be directly related to the ablated internal structure of the networks. That is, if an ablated area had little or no functional localization, then little or no behavioral dissociations were observed. If, however, the ablated area had very specific internal structure, then very specific behavioral dissociations were observed. It is important to note, however, that there was not a one-to-one correspondence between internal structure and behavioral dissociations, implying that cognitive neuroscientists must be careful when using lesioning data to theorize about the functional architecture of cognition.