This study examined the effects of encoding instructions and subjects' prior knowledge on sensitivity to feature co‐occurrence in the learning of two novel, ill‐defined categories. Sixty‐five subjects were allocated to conditions in which they were either required to discriminate between the members of the categories (intentional instructions) or to identify the features of each category instance (incidental instructions). Categories were presented with neutral labels or labels that cued the retrieval of prior knowledge. This additional information was presented at the beginning or the end of the study phase. All instances were presented four times, followed by a test phase in which subjects classified old and new stimuli that varied in the extent to which they preserved the correlations between feature values that were present in the study items. Analyses of test‐phase categorisation indicated that intentional encoding led to greater sensitivity to feature correlations than incidental encoding, and that provision of meaningful category labels facilitated the encoding of correlations between features that were seen to be consistent with prior knowledge. The implications of these results for models of the interaction between similarity‐based processes and prior knowledge are