According to authors, these poor results are not deficiencies in the software per se but instead a mismatch between how we remember faces and how composites are produced. "Numerous lines of evidence converge on the view that faces are generally processed, stored and retrieved at a holistic level rather than at the level of individual facial features." Ultimately the psychological process of remembering faces may include more complex representations such as multidimensional similarity to other faces or relative sizes and distances of features and so on that are not readily retrieved by memory nor utilized by facial composite software.The authors go on to recommend whole face, or "looks like", methods instead of the existing "parts" based methods (e.g. eyes and ears) as a way to get better results. "Looks like" systems have produced good results in previous studies, as I recall, so maybe they are on to something.
The study was published in Current Directions in Psychological Science (February 2007 issue).