Recent cognitive theories propose that attentional biases cause or maintain anxiety disorders. This study had several aims: (i) to investigate such biases in generalized anxiety disorder (GAD) using naturalistic, ecologically valid stimuli, namely, emotional facial expressions; (ii) to test the emotionality hypothesis by examining biases for happy as well as threat faces; and (iii) to assess the time course of the attentional bias. The dependent variable was an index of attentional bias derived from manual RTs to probe stimuli. There were four independent variables: one between-subjects variable of group (2: GAD, control), and three within-subjects variables: Type of emotional face (2: threat, happy), Stimulus duration (2: 500 ms, 1250 ms) and Half of task (2: first, second). Attentional bias was assessed with a dot probe task. The stimuli comprised photographs of threatening, happy and neutral faces, presented using two exposure durations: 500 ms and 1250 ms. Anxious patients showed greater vigilance for threatening faces relative to neutral faces, compared with normal controls. This effect did not significantly vary as a function of stimulus duration. Anxious patients also showed enhanced vigilance for happy faces, but this was only significant in the second half of the task. The study confirmed not only that GAD patients show a bias in selective attention to threat, relative to controls, but also that this bias operates for naturalistic, non-verbal stimuli. As the attentional biases for threat and happy faces appeared to develop over a different time frame, different underlying mechanisms may be responsible.