We explore whether social context affects how labels (relative to other features) affect category learning. We taught 104 participants four novel categories using a feature inference task. In a between-participants design, we manipulated: 1) the social context of the task (social context vs. on the computer); and 2) which dimension of the category members could be used to perfectly predict the target feature: the category label, a biased feature (which is salient and already associated with the target feature in the correct way) or a non-biased feature (which is less salient and not already associated with the target feature in any way). Learning curves were used to assess whether participants assumed that labels were uniquely helpful compared to other features. The results suggest that the extent to which labels are privileged depends on the context in which the category learning task is presented. When the task is social, people learn quickly regardless of whether a label or another feature is the most informative. When the task is not, both novel labels and biased features are more useful than non-biased features.