The extent to which human discrimination learning is based on elemental or configural stimulus representations was examined in 7 experiments. In Experiments 1a and 1b, participants were able to learn nonlinear discrimination problems in a food-allergy task. In unique-cue theories, such learning is explained by individual stimulus elements acquiring independent connections with the outcome and also combining to form unique cues that function elementally. In Stage 1 of Experiments 2, 3, and 4a-c, Food A signaled an allergy outcome (O) (A → O) when presented alone but signaled no allergy (AB → no O) when paired with Food B. In Stage 2, Food B was paired with the allergy (B → O). In a test phase, the original discrimination between A and AB was found to be intact, at variance with the unique-cue theory. By contrast, in Experiments 5a, 5b, and 6, an effect of the B → O trials on the A-AB discrimination was observed with training procedures previously found by D. A. Williams (1995) to encourage elemental processing. Experiment 7 showed that the influence of B → O trials on the A-AB discrimination was unaffected by pretreatments designed to foster an elemental processing strategy. Overall, these results are problematic for unique-cue theories and instead imply that, by default, configurations of elements form the representational basis of simple discrimination learning in humans.