How do the ways in which we learn influence our cognitive representations of what we learn? We show that in language learning tasks, the process of learning to conceptualize and categorize perceptual input shapes how that input gets represented in mind. In representation, there seems to be a give and take between veridicality and completeness, on the one hand, and discrimination and accurate categorization, on the other. Learning to better discriminate objects into categories based on their highly-salient features makes people less likely to notice or remember the same objects' less-salient features. Gains in response-discrimination between categories thus come at a cost to within-category discrimination. This is a natural consequence of error-driven learning, a mechanism underlying most contemporary learning models. We present an exposition of error-driven learning, outline its implications for cognitive representations, and test these predictions, showing that the patterns of human learning are consistent with our analysis. We suggest that the mechanisms of human learning obey a simple principle: there can be no representation without taxation.