We use a masked priming procedure to test two accounts of the picture-word interference (PWI) effect: the lexical selection by competition account (Levelt et al., 1999; Roelofs, 1992) and the response selection account (Lupker, 1979; Miozzo and Caramazza, 2003). In the visible (standard) condition, we replicated the often-observed semantic interference effect. In the masked condition, we observed semantic facilitation. We take the polarity shift as a function of masking to mean that the semantic interference and semantic facilitation in the PWI task should be attributed to two qualitatively different processes. We argue that this conclusion follows naturally from the response selection account, but only with great difficulty from the lexical selection by competition account.