This paper investigates false recognition of semantically related distractor items. Using the Deese (1959) false memory paradigm, we investigated the predictions that MINERVA2 (Hintzman, 1988) makes for true (exemplar) and false (prototype) recognition. In MINERVA2, exemplar recognition is primarily due to a high degree of similarity to a single memory trace and prototype recognition is due to the summing of smaller amounts of similarity across many memory traces. MINERVA2 predicts that: (1) while performance for exemplars and prototypes increases at approximately the same rate at low levels of learning, additional learning produces a greater increase in performance for exemplars than for prototypes; (2) reducing the number of exemplars studied will decrease performance more for prototypes than for exemplars; and (3) decreasing associative strength of an exemplar to a prototype should increase exemplar performance while decreasing prototype performance. The results of four experiments support the predictions of the model and are consistent with the notion that true recognition is primarily due to a high degree of match to a single trace, while false recognition is due to the summing of smaller amounts of similarity to numerous traces.