In two experiments, participants studied two types of word lists. Direct lists were taken from the Deese-Roediger-McDermott (DRM) paradigm (e.g., "water", "bridge", "run") and contained words directly related to a nonpresented critical item (CI; e.g., "river", Roediger & McDermott, 1995). Mediated lists (e.g., "faucet", "London", "jog") contained words related to the CI through a nonpresented mediator. After each study list, participants completed either a recall test, a recall test with a warning about the CI, arithmetic problems, or a recognition test, or they guessed the CI. On a final recognition test, both warning and guessing decreased direct false recognition but increased mediated false recognition, an ironic effect of guessing. An initial recognition test also increased final mediated false recognition. We argue that warning and guessing tasks strengthened associative pathways to the CI, increased the accessibility of associated mediators, and increased monitoring for the CI at test. Increased monitoring was able to reduce CIs from direct, but not mediated, lists. (Contains 5 tables and 4 footnotes.)