A. H. Criss and R. M. Shiffrin (2004) argued against the composite context noise explanation of recognition memory introduced by Dennis and Humphreys (2001) by showing that with novel face stimuli, related distractors that are drawn from study categories show elevated false alarm rates. Dennis and Humphreys (2001) proposed two mechanisms by which related false alarm rates might arise. The first mechanism posited that such items might be produced as implicit associative responses. Such a mechanism is unlikely to apply to face stimuli. The second mechanism posited a category wide criterion shift, which Criss and Shiffrin (2004) argued was implausible because it requires criterion to be adjusted on an item by item basis during test. Rather they contended that direct interference between study items was more likely to account for the data and fit the Retrieving Effectively from Memory model Shiffrin and Steyvers (1997) to their data. We suggest that the mechanism by which false alarms are generated for word and novel face stimuli may differ. Instead of focusing on related distractors, we focus on unrelated distractors. Using words, we show that if list length is manipulated by keeping the number of categories constant but increasing the number of exemplars in each category, then the unrelated false alarm rate decreases - thus inducing an inverse list length effect in which performance on short lists is worse than on long lists. Simulations demonstrate that exemplar models such as REM are not capable of accounting for the results without modification. Rather a composite representation of the studied categories that subjects can use to reject unrelated lures must be formed.