Matching two different images of a face is a very easy task for familiar viewers, but much harder for unfamiliar viewers. Despite this, use of photo-ID is widespread, and people appear not to know how unreliable it is. We present a series of experiments investigating bias both when performing a matching task and when predicting other people’s performance. Participants saw pairs of faces and were asked to make a same/different judgement, after which they were asked to predict how well other people, unfamiliar with these faces, would perform. In four experiments we show different groups of participants familiar and unfamiliar faces, manipulating this in different ways: celebrities in experiments 1-3 and personally familiar faces in experiment 4. The results consistently show that people match images of familiar faces more accurately than unfamiliar faces. However, people also reliably predict that the faces they themselves know will be more accurately matched by different viewers. This bias is discussed in the context of current theoretical debates about face recognition, and we suggest that it may underlie the continued use of photo-ID, despite the availability of evidence about its unreliability.