Dennis and Humphreys (2001) proposed that interference in recognition memory arises solely from the prior contexts of the test word: Interference does not arise from memory traces of other words (from events prior to the study list or on the study list, and regardless of similarity to the test item). We evaluate this model using output interference, a decline in accuracy as a function of the words presented during test. Output interference is consistent with models that allow interference from words other than the test word, when each test produces a memory trace, and hence a source of interference. Models positing interference solely from prior contexts of the test word itself predict no effect of items presented during test, without added assumptions. We find robust output interference effects in recognition memory. The effect remains intact after a long delay, when study-test lag is held constant, when feedback is provided, and when the test is yes/no or forced choice. These results are consistent with, and support the view that interference in recognition memory is due in part to interference from words other than the current test word.