There is now ample evidence that extinction, the loss of learned performance that occurs when a Pavlovian signal or an instrumental action is repeatedly presented without its reinforcer, does not reflect a destruction of the original learning. This article summarizes the evidence and extends and updates earlier reviews. The main alternative to "unlearning" is the idea that extinction (as well as other retroactive interference processes, including counterconditioning) involves new learning that is stored along with the old. One consequence is that the Pavlovian signal or instrumental action has two available "meanings" and thus has the properties of an ambiguous word: its current meaning (and the resulting behavioral output) depends on what the current context retrieves. Contexts can be provided by a variety of background stimuli, including the physical environment, internal drug state, and time. The second thing learned (e.g., extinction, counterconditioning) seems especially dependent on the context for retrieval. A variety of evidence is consistent with this analysis, which highlights several important sources of relapse after extinction. The article concludes with several issues for future research, among them the question of how we can optimize extinction and other putative "unlearning" treatments so as to prevent the various forms of relapse discussed here.