Deep encoding, relative to shallow encoding, has been shown to increase the probability of false memories in the Deese/Roediger-McDermott (DRM) paradigm (Thapar & McDermott, 2001; Toglia, Neuschatz, & Goodwin, 1999). In two experiments, we showed important limitations on the generalizability of this phenomenon; these limitations are clearly predicted by existing theories regarding the mechanisms underlying such false memories (e.g., Roediger, Watson, McDermott, & Gallo, 2001). Specifically, asking subjects to attend to phonological relations among lists of phonologically associated words (e.g.,weep, steep, etc.) increased the likelihood of false recall (Experiment 1) and false recognition (Experiment 2) of a related, nonpresented associate (e.g.,sleep), relative to a condition in which subjects attended to meaningful relations among the words. These findings occurred along with a replication of prior findings (i.e., a semantic encoding task, relative to a phonological encoding task, enhanced the likelihood of false memory arising from a list of semantically associated words), and they place important constraints on theoretical explanations of false memory.