The present experiments examined the extent to which two possible sources of error affect healthy subjects' performance in a rule-shift task. All 115 participants first received a discrimination learning task, in which a pair of different visual stimuli was presented on each trial, one of which had to be identified as 'correct.' Each stimulus varied in two dimensions: a task-relevant and a task-irrelevant dimension. Feedback on correctness was given after each choice. After eight successive correct choices, the nature of the task-relevant dimension changed: the post-shift learning phase. Two types of error can occur in this phase: continued responding to the former relevant, but now irrelevant, dimension, a perseverative error, and non-responding to the former irrelevant, but now relevant, dimension, an error due to learned irrelevance. Different groups received a post-shift task in which none, one, or both of these two types of error could affect performance. The number of incorrect choices in the post-shift phase was significantly affected by learned-irrelevance errors but not by perseverative errors. An associative-learning model incorporating feedback-induced changes in both associative strength and saliency of the elements comprising the stimuli can explain these results.