Many birds and mammals respond to a heightened risk of predation, especially that associated with smaller group sizes, with an increase in vigilance. All interpretations of the way in which vigilance responds to changes in predation risk assume that animals feeding with their heads down (i.e. animals in a nonvigilant state) cannot detect approaching predators. We provide the first explicit test of this assumption by 'flying' a mounted hawk down a 15-m chute towards actively feeding, free-living, dark-eyed juncos, Junco hyemalis. Juncos were targeted individually for simulated attack when they had either a 'head-down' view up the chute, or a completely unobstructed view; a junco with a head-down view could see up the chute only when it lowered its head to feed. Juncos with an unobstructed view almost always detected the hawk at the maximum distance of 15 m. Juncos with a head-down view usually detected the attack at a distance of 10-15 m against a grey background, but detection distances were shorter when attacks occurred against a camouflaged background. The results demonstrate that these birds have a considerable ability to detect approaching predators even when not overtly vigilant, although their detection ability is greater when they raise their heads. Vigilance sequences, therefore, probably consist of bouts of low-quality detection (active feeding) interspersed with bouts of higher-quality detection (overt vigilance) that can only be accomplished at the expense of feeding. This realization has major implications for current interpretations of the vigilance group size effect and antipredator vigilance in general. Copyright 1999 The Association for the Study of Animal Behaviour.