For a growing number of applications it is important that the perceptual behavior of a virtual human be at once realistic and believable. To be realistic, a model needs to be psychologically plausible, and to be believable the behavior must look right. To this end, I describe a model of attention that was implemented in a virtual helicopter pilot using the Soar cognitive architecture. It is based on psychological principles and where the pilot performs a number of realistic tasks in a real-time, synthetic environment. We are now extending the model to be more believable by incorporating gaze behaviors where the virtual human has a head and eyes. This paper describes the basic model of attention and the first steps toward achieving more believable gaze behaviors.