Short-latency saccades to targets among nontarget backgrounds are often directed to the center of the entire (target + nontarget) stimulus configuration. This "averaging" or "center-of-gravity" tendency has been attributed to an automatic, reflexive saccadic response to a poorly-resolved visual signal. We investigated the role of high-level processes by varying the probability of the target appearing in one of two locations. Subjects were asked to make a saccade to a target "+" located above-right or above-left of a central fixation point. A nontarget ("x") was in the other location (directional separation = 30 deg). The mean latencies were short (180-230 msec) in accordance with instructions. Mean saccadic direction was shifted to the right by 24-52% of the directional separation of the stimulus pair as the probability of the target appearing on the right increased from 0.2 to 0.8. The difference in saccadic directions as a function of the actual target location was small and independent of probability, showing that probability introduced a bias without affecting the discriminability of the target from the nontarget. The effect of probability was reduced when the discrimination of the target from the nontarget was easier (square vs triangle), and abolished (saccadic accuracy near perfect with the same average latencies) when the target was presented alone. The results show that the direction of short-latency saccades, initiated before the target has been distinguished from a nearby nontarget, is based on the prior history of target locations and expectations about the future location of the target. High-level plans can account for effects of nontargets on saccades. To infer that a reflexive sensorimotor averaging mechanism exists solely on the basis of observed saccadic "centering" tendencies is unwarranted.