In making sense of the environment, we implicitly learn to associate stimulus attributes that frequently occur together. Is such learning favored for categories over individual items? Here, we introduce a novel paradigm for directly comparing category- to item-level learning. In a category-level experiment, even numbers (2,4,6,8) had a high-probability of appearing in blue, and odd numbers (3,5,7,9) in yellow. Associative learning was measured by the relative performance on trials with low-probability (p = .09) to high-probability (p = .91) number colors. There was strong evidence for associative learning: low-probability performance was impaired (40ms RT increase and 8.3% accuracy decrease relative to high-probability). This was not the case in an item-level experiment with a different group of participants, in which high-probability colors were non-categorically assigned (blue: 2,3,6,7; yellow: 4,5,8,9; 9ms RT increase and 1.5% accuracy increase). The categorical advantage was upheld in an explicit color association report (83% accuracy vs. 43% at the item-level). These results support a conceptual view of perception and suggest empirical bases of categorical, not item-level, color labeling of learning materials.