An experiment involving 308 human subjects compared the difficulties of four types of shift learning. Initial learning was of an exclusive-or (XOR) structure on 2 of 3 stimulus dimensions. One shift type was a reversal, a second was to a single previously relevant dimensions, a third was to a single previously irrelevant dimension, and a fourth was to an XOR on 1 previously relevant dimension and 1 previously irrelevant dimension. The resulting ease of shift learning was in the same order. The results cannot be fit by some recent connectionist models of category learning, despite modifications to those models that allow them to show reversal-shift advantage in simpler situations, nor can the results be fit by some rule-based theories. The results also appear to conflict with earlier findings that a shift from an XOR to another XOR can be easier than a shift from an XOR to a single dimension, but a reconciliation is found.