When crime-fighting tools go bad: Problems with the face-composite system

i-8642feb8af0b92e67c695bafb9f721fc-face.jpg

i-eca0cf2af9fc3ac4445c7dff7d8aab70-research.gifThousands of police departments use face composite software to help create a picture of crime suspects. You've probably seen one of the systems in use on TV: witnesses build a picture of the suspect by choosing each individual facial feature -- hair, eyes, nose, and so on. But what happens when the suspect is captured and the witness is asked to identify the real perpetrator in a lineup? Does the witness remember the actual face they saw at the crime scene, or the composite face created at the police station? A recent study has found that the process of creating a face composite can have a dramatic impact on the memory of a real face.

Research in face perception shows that we tend to remember faces not so much based on the individual features, the way face composite software works, but based on the relationship between the elements of a face. It makes sense: after all, we can still recognize our friends if they get a haircut or shave off their mustache. A team led by Gary Wells reasoned that if crime witnesses spend 20 minutes recreating a face they saw for perhaps only a few seconds at the crime scene, the face they create might supplant the real thing in memory. While a few studies have tested this notion previously, they were typically done on too small of a scale to have conclusive results. In an effort to find a more definitive answer, Wells' team created two studies with over 300 participants.

In the first experiment, 150 participants were divided into 3 groups. All the "witnesses" were told they were going to be participating in a study on "perceptions of people." They were first shown one of 50 different photos and asked to rate it on 10 different traits, such as intelligent, humorous, and so on. Next the picture was removed and they were asked to write a description of the face. The first (control) group was sent home and asked to return 48 hours later. The second group was asked to use FACES computer software, the most popular face composite software, to create a composite image of the face they had seen. The third group didn't create the composite, but was instead shown a composite of the face they had seen, created by one of the participants in the second group.

Finally, in 48 hours, all the participants returned and were asked to pick the photo they had previously seen out of a lineup of six photos, which had been selected for similar general characteristics to the "suspect" (this is standard police procedure -- it wouldn't be fair, for example, to place a black suspect in a lineup of white people, or a bearded suspect in a group of clean shaven people). They were initially told that the suspect may or may not be in the lineup, and so the correct answer might be "none of the above" (again, standard police procedure). If no suspect was selected at this point, then participants were asked who they would pick if they had to make a choice. Here are the results:

i-36ed8d63394760d49c5527be52a44a53-face2.gif

As you can see, the results are dramatic. Just 10 percent of the people who built composites correctly identified the suspect initially. Even when forced to make a choice, their accuracy only climbed to 30 percent, not much better than chance. By contrast, the control group, who had only seen the original photo, was 84 percent accurate originally, and 94 percent when forced.

But these results aren't ideal. Witnesses see real crimes, not static photos. And misidentifications when the perpetrator is in the lineup aren't catastrophic, because the "filler" lineup participants are not going to be prosecuted -- the problem comes when the police arrest the wrong suspect and he or she is picked out of a lineup. In the second experiment, Wells' team addressed these issues in a new experiment by showing participants a simulated crime video, and by using some lineups that didn't actually include the perpetrator.

As before, the 200 "witnesses" were told that they were in an experiment on "perceptions of people." They were told to watch the crime video closely because they'd be asked questions about it later. When the video was over, they were told that the video showed a person planting a bomb in a building's air shaft, and they would need to write down a description of the "roof bomber," whose face had been clearly visible in the video for 21 seconds. Half the participants were sent home, and the other half made composites, as before.

48 hours later, all participants were called back and asked to identify the perpetrator in a lineup, exactly as before. This time, however, half the lineups didn't included the perpetrator at all. Here are the results:

i-84bc54717ca8297b1b3a1ac4c41e0b9f-face3.gif

There was no difference between the lineups with and without the perpetrator, so that data is not presented here. As you can see, however, those who made composites made many fewer correct IDs when the perpetrator was present, and also selected the incorrect suspect significantly more often, even when the perpetrator was not present in the lineup. In this more realistic scenario, creating a composite face still puts the witness at a whopping disadvantage compared to people who had only described the ciminal verbally.

Wells' team is careful to point out that face composites can still be an important crime-fighting tool. If there are several witnesses, one could make the composite and the others could ID the suspect in a lineup. A face composite may be the only way police have of locating the suspect at all; when captured, other evidence might link the individual to the crime.

However, Wells et al. do point to an older technology which may need to be revisited: the police sketch artist. Perhaps if a sketch artist creates an image based on a witness's description instead of relying on composite-creating software, the witness may still accurately remember the perpetrator.

Wells, G.L., Charman, S.D., & Olson, E.A. (2005). Building face composites can harm lineup identification performance. Journal of Experimental Psychology: Applied, 11(3), 147-156.

More like this

Face-composite software is commonly used to generate images of crime suspects. But how accurate is it? We've reported here on a study suggesting that building face-composites may actually harm the memory of eyewitnesses. Now a new review article is suggesting that there are additional problems with…
Click on the image below to be taken to a quicktime movie showing 9 different faces. When the movie is finished playing, drag the slider back and forth to pick the face you think is the most attractive. The faces are composite images—"average" faces made by morphing together 48 different photos…
In every courtroom drama, the most dramatic scene is always when the star witness points her finger at the villain and proclaims that "he did it!" The confidence with which an eyewitness describes the perpetrator of a crime is often the most convincing evidence in a court battle. But how accurate…
I never really understood how anyone could identify someone off of those police sketches! Now it seems that some psychologists have come to the same conclusion - but not pulling reasons out of their rear ends...like me! In an article appearing in the February issue of Current Directions in…

Good article. It's scary how many innocent people could be languishing in jail just based on eyewitness testimony. Police sketch artists pose their own problems, too (the skill of the artist must be taken into account). And then there's the problem of prosopagnosia, which is fairly common. As for my own experience in working in a convenience store, I have to ID people. Frequently I get, "Don't you recognize me? I'm in here all the time!" Well, gee, I get a hundred customers a night and unless someone has very distinctive features or comes in every night, then no, I don't recognize you. Some years ago, while working in a video store, I had an individual run out the the store with a stack of videos. I thought I could finger the guy, but the photo line-up I was shown consisted of a bunch of faces of guys who could have been brothers. Unfortunately, other descriptive factors such as height, weight, etc. were not included.

By David Group (not verified) on 27 Sep 2006 #permalink

This exactly matches my completely ignorant expectation. I've always thought face composite contruction would lead to this sort of result.

...which doesn't mean a damn thing, but it's always gratifying to have your world view confirmed by real science. :

By SmellyTerror (not verified) on 28 Sep 2006 #permalink

yeah great article!

And then there's the thoroughly discredited police technology of polygraphs. I still see mentions in the news about this or that suspect passing or failing a police lie detector test.

Last I heard the FBI still use it as a pre-employment screening test. It is also apparently still popular with the CIA.

The only valid results occur when a naive subject is intimidated into confessing something. Anyone who has been rehearsed can easily force the results in his favour.

With such ambiguous recordings, it is easy for some operator/interrogators to rely on their own 'intuition' or prejudices in picking an outcome. Since a bullying approach is part of their training, subjects who object to this treatment can easily be punished with a failing result. There can be no effective supervision against vindictive operators,since everything is subjective.

Many careers have been ruined, innocents convicted and the guilty set free because of reliance on this fraudulent technology.

Reference: www.antipolygraph.org

By JohnnieCanuck (not verified) on 02 Oct 2006 #permalink

Yes, prosopagnosia is just one of the many sensory deficits in cognition. The studies I did at the Open University on such conditions post graduate with behaviour sciences led me to understand my own and others semantic errors. I give this link which is good in detailing the sheep experiment.

http://www.acnr.co.uk/pdfs/volume4issue5/v4i5cognitive.pdf#search=%22co…

A farmer with prosopagnosia (there are a range of agnosias), could remember and tell by actually looking at his own sheep and at photos of them. They had a 'value'. In looking at known public figures as photographs he could not identify the name or who that person was, or what he did. Some issue made of facial mapping software which we used that had missing nodes as comparative. It guessed the missing features from its database. (Neurotool)was the programme used as a network system of facial mapping.

In contrast a human has sensory inputs in sight, sound, touch, feel, and smell of an organic object. If one sees something familiar as trained then it is remembered with some association. But creating lab type conditions in monitoring subjects has problems in setting the scene empirically, that is it is contrived and at set times of no conflict with external and internal conflicts.

For example a witness may see part of a face, an outline, and a expression which a computer cannot simulate in facial recognition databases. Yet the police artist is listening to what is said and using his or her own sensory experience to draw a feature and keep changing it uniquely.

I agree that the interviewer has biases and brings his or her own values on the outcome as perceptive. However, the skill of the artist is a factor too as intuitive to the person who describes a feature. Meanings of features are different in description as the sheep farmer case highlights.

Facial composites take away the very essence of expression. In the undergraduate years we were told to turn down the sound, and interpret/observe the body language and delivery of a person on television to assess the hidden meanings. In doing so since then I have learned that a 'face' has a personality that changes that facial movement with each emotion. That is how actors and actresses change their face.
It is now called in US 'thoughtprints' in statement analysis. A face has feelings subconsciously.

Consider also Wernicke-Korsakoffe complications and blindsight as other conditions undiagnosed in the witness or disclosed. I am both aphasic and dyslexic in aquired dyslexia, but have a coping mechanism of remembering signs, codes and images. In contrast common words are a problem to me, but medical terminology is my own. Same as the farmers in interest value and emotion. So if we like a 'face' do we remember it more than one we do not?

In closing I therefore support the article from a multidisciplined perspective.

By Pauline Anne Badger (not verified) on 08 Oct 2006 #permalink