In military operations, quick and accurate target detection and identification is critical for mission success. Augmented reality (AR) technologies can aid target detection and identification by layering digital imagery atop a Soldier's field of view to increase situational awareness. These systems are rarely perfect, however, and in some cases unreliable AR may actually interfere with performance. This investigation focused on the capacity for unreliable AR to impair performance. We showed participants a series of 2D simulations where highly-visible AR cues were superimposed over tanks placed randomly in a grassland environment. The reliability of these cues varied (from 25% to 100%) throughout the experimental session, as some valid targets were erroneously un-marked (false negatives) while some invalid targets were erroneously marked (false positives). Participants were asked to search for the vehicles while being assisted by the AR; search accuracy and response time were analyzed, and participants provided feedback regarding their mental workload and trust in the AR. We found the expected negative relationship between unreliability and performance, but also found that AR false positives were more damaging to performance than AR false negatives. Unreliable AR also hurt performance more when marking vehicles at greater distances. Further, although error type and target distance had powerful effects on participant performance, they had less of an impact on subjective trust and workload, suggesting that Soldiers using AR might not be consciously aware of how their own performance changes as a function of AR properties. In summary, unreliable AR hurt performance differently depending on the type of errors produced by the system, and impaired some aspects of performance but not others. These results carry important implications for how AR is designed to improve performance on the battlefield. Keywords: augmented reality, trust in automation, visual detection and identification, mental workload, human performance