The human amygdala responds selectively to consciously and unconsciously perceived facial expressions, particularly those that convey potential threat such as fear and anger. In many social situations, multiple faces with varying expressions confront observers yet little is known about the neural mechanisms involved in encoding several faces simultaneously. Here we used event-related fMRI to measure neural activity in pre-defined regions of interest as participants searched multi-face arrays for a designated target expression (fearful or happy). We conducted separate analyses to examine activations associated with each of the four multi-face arrays independent of target expression (stimulus-driven effects), and activations arising from the search for each of the target expressions, independent of the display type (strategic effects). Comparisons across display types, reflecting stimulus-driven influences on visual search, revealed activity in the amygdala and superior temporal sulcus (STS). By contrast, strategic demands of the task did not modulate activity in either the amygdala or STS. These results imply an interactive threat-detection system involving several neural regions. Crucially, activity in the amygdala increased significantly when participants correctly detected the target expression, compared with trials in which the identical target was missed, suggesting that the amygdala has a limited capacity for extracting affective facial expressions.