Artificial face recognition systems typically do not attempt to handle very variable images. By comparison, human perceivers can recognize familiar faces over much more varied conditions. We describe a prototype face representation based on simple image-averaging. We have argued that this forms a good candidate for understanding human face perception. Here we examine the stability of these representations by asking (i) how quickly they converge; and (ii) how resistant they are to contamination due to previous misidentifications. We conclude that face averages provide promising representations for use in artificial recognition.