Visual working memory (VWM) is characterized as extremely capacity limited. This finding is supported by the dramatic decline in change detection performance beyond a small number of items, as well as the observation of flat error distributions in delayed-estimation tasks. However, continuous resource models predict that small amounts of memory resources can be distributed to items at the expense of memory resolution (resulting in low response precision). These low-resolution memory representations should have nearly flat error distributions that could appear indistinguishable from uniform guessing distributions. In the current study, memory resource allocation was manipulated by varying the probability of an item being probed at recall. Responses were intermixed between continuous response and two-alternative forced-choice (2AFC) trials to examine whether these low-probability items could produce above-chance performance, consistent with them being held in memory. For comparison with the distribution of continuous responses, the magnitude of the discrimination between the target and lure colors was manipulated. Accuracy on the 2AFC trials was sensitive to both discrimination difficulty and probe probability manipulations. Critically, above-chance performance was found in the lowest probe probability condition (10% probe probability, equivalent to an item load of 10), suggesting that this condition had low-resolution memory representations rather than no memory representations. These findings are consistent with the predictions of continuous resource models and applications of signal detection models of VWM.