Models of category learning can take two different approaches to representing the relationship between objects and categories. The generative approach solves the categorization problem by building a probabilistic model of each category and using Bayes' rule to infer category labels. In contrast, the discriminative approach directly learns a mapping between inputs and category labels. With this distinction in mind, we revisit a previously studied categorization experiment that showed people are biased towards categorizing objects into a category with higher variability. Modelling results predict that generative learners should be more greatly affected by category variability than discriminative learners. We show that humans can be prompted to adopt either a generative or discriminative approach to learning the same input, resulting in the predicted effect on use of category variability.