We argue that are no such things as literal categories in human cognition. Instead, we argue that there are merely temporary coalescences of dimensions of similarity, which are brought together by context in order to create the similarity structure in mental representations appropriate for the task at hand. Fodor (2000) contends that context-sensitive cognition cannot be realised by current computational theories of mind. We address this challenge by describing a simple computational implementation that exhibits internal knowledge representations whose similarity structure alters fluidly depending on context. We explicate the processing properties that support this function and illustrate with two more complex models, one applied to the development of semantic knowledge (Rogers and McClelland, 2004), the second to the processing of simple metaphorical comparisons (Thomas and Mareschal, 2001). The models firstly demonstrate how phenomena that seem problematic for literal categorisation (such as the ‘non-literal’ comparisons involved in metaphor and analogy) resolve to particular cases of the contextual modulation of mental representations; and secondly prompt a new perspective on the relation between language and thought: language affords the strategic control of context on semantic knowledge, allowing information to be brought to bear in a given situation that might otherwise not be available to influence processing. This may explain one way in which human thought is creative, and distinctive from animal cognition.