When two complex natural scenes are simultaneously flashed left and right of fixation, subjects can make reliable saccades toward animal targets in as little as 120 ms (Kirchner & Thorpe, Vis Res, 2006). This is less than half the time required to perform a manual go/nogo task with the same sort of material (Bacon-Macé et al, JEP:HPP in press). One feature of manual tasks is that participants can easily switch between target categories from block to block, for example, between animals and means of transport (VanRullen & Thorpe, Perception, 2001), or between animal and human faces (Rousselet et al., JOV, 2004). Here we tested whether subjects can also easily reverse the target category in a choice saccade task. To our surprise, we found very strong asymmetries. For example, when using photographs of humans and means of transport, we found that when humans were the target, saccades were initiated in as little as 100 ms with mean RT of 140 ms and accuracy of 89%. However, using the same materiel, when the target was a means of transport, performance was much poorer (mean RT = 170 ms, 71% correct). Furthermore, fast saccades were often made in the wrong direction as if the eye is automatically drawn to one of the two stimuli. Since there are populations of neurons that fire earlier to human faces than to other classes of stimuli (Kiani et al, J Neurophysiol, 2005), such neurons could feed into a form of high-level saliency map and thus lead to very fast saccadic reactions to important types of stimulus. However, unlike slower manual responses, it may be that this sort of activation cannot be modulated by top down control, making it difficult or maybe even impossible to modify the target stimuli category from block to block.