Often the test is done with words. A subject will be shown a series of words on a computer screen, flashed in rapid succession, and they'll have to identify certain target words (Cognitive Daily has an example, go check it out). Distractor words are placed in the sample in order to measure the effect those types of words have on the task.
This molester study was carried out with pictures. In a series of rapidly flashed pictures, subjects had to try to identify target pictures (chairs and trains), and two types of distractors were used, pictures of children and of animals. The researchers found a significant difference between the accuracy rates of non-molestors and convicted child molestors. In other words, molestors are more distracted in this task by pictures of children.
So would this be a reliable way of identifying child molesters? I don't think so. Dave Munger points out:
Others who are interested in kids, like teachers, new parents, and grandparents, might also be distracted by pictures of kids. They've done some preliminary research suggesting this is not the case, but perhaps other non-dangerous populations would be mistakenly identified using this test.
As I noted in the comments over there, one of the first things I thought about were parents who had lost a child. They would be a relatively small segment of the population, but my guess is that you would see a large distraction effect for pictures of children in this group. And you really wouldn't want to tag a parent who had lost a child as a potential child molester.
That's one big problem with psychological tests that measure a specific response. If you don't have a firm grasp on the mechanism at work, and all possible causes of the effect, then it's far too blunt an instrument to really do anything useful. In this particular case, the risk of false positives is so great that you really wouldn't want this approach anywhere near a police station or a court of law.