Recent empirical studies of semi-supervised category learning—where learners only occasionally receive information about a given item’s category membership—have yielded contradictory results, with some studies showing strong effects of unlabeled experience and others little or no effect. We report two experiments designed to help understand this heterogeneity. In both, participants performed a two-category classification task with novel stimuli varying along two psychologically separable dimensions. In semi-supervised conditions, participants categorized and received feedback on 32 “labeled” items intermixed with a large number of “unlabeled” items. In the supervised-only condition, participants viewed the same labeled trials intermixed with a large number of filler trials. Without time pressure participants learned the task equally well in both conditions. When required to respond very rapidly, however, participants performed substantially better in the semi-supervised condition. The discrepant results may indicate a role for selective attention in human semi-supervised learning.