Patients with Mild Cognitive Impairment (MCI) present with feature binding impairment (Koppara et al., 2015). Such impairment has often been greater than that found when they retain individual features. Patients with Alzheimer's disease (AD) have typically and consistently shown this dissociation (Parra et al., 2009; 2010; 2011), suggesting that binding functions carried out in short-term memory (STM) are selectively affected by the disease. However, findings have recently become controversial indicating that such a dissociation (i.e., binding deficits >> feature processing deficits) may be contingent upon memory load. It is important to demonstrate the exact setting with which such a selective binding impairment can be identified. This study addressed such a hypothesis. Two groups of participants (Healthy Older Adults –HOA- and amnestic MCI patients) were assessed with a standard neuropsychological battery and two versions of the STM binding test. In one version participants studied visual arrays of three items which were either random shapes (single features) or colored shapes (feature binding). They then decided if arrays that followed were the same or different. In the other version the arrays consisted of two items. We used a 2x2x2 ANOVA with Condition (shape vs binding), Setting (3 items vs 2 items) and Group (HOA vs aMCI) as the investigated factors. We found a significant three-way interaction [F(1,180) = 21.03, p ; 0.001, h2=0.105, b=1.0]. Performing the STM task with two items yielded the expected dissociation whereby aMCI were found to be significantly poorer than HOA during the binding than during the shape only condition. Such dissociation disappeared with 3 items. Interestingly, aMCI who performed the 2-item version of the task were more cognitively impaired (lower MMSE) than those who performed the 3-item version suggesting that this dissociation persists even in the advanced stages of aMCI but only when STM is not overloaded. It is the feature binding condition of the STM binding test which has proved sensitive and specific to AD. Such diagnostic value has been highlighted by the selectivity of this impairment. Our results suggest that if such selectivity is desired, memory load should be controlled for. Mean proportion of correct recognition across the two memory load settings, experimental conditions, and groups.