One approach to detecting failures in dynamic systems is to run a model in parallel to the real system. Tn such cases the operator is expccted to make decisions about possible failures by comparing the outputs of the system and its model. This task is discussed from the point of view of human information processing. An experiment was run to test different means for aiding the operator in this task. It was found that the detection proccss may be improved by providing the operator with estimates of the failure odds. Moreover, when both the system and the model outputs were noisy, additional improvement was achieved by providing the operator with pre- smoothed outputs.