When developing new software for polygraphs (lie detectors) professionals have to face one very obvious problem. The fact that it is created in an artificial environment where to achieve the effect of scientists have to lie specifically, the machine eventually be able to understand the issue. However, laboratory conditions and reality differ dramatically. In reality, the lie nobody simulates, and therefore the machine can be easily fooled if you know how to do it.
In order to create a very good lie detector, it is necessary to train machine based on real examples of lies. That’s why scientists from the University of Michigan (USA) decided to go to the different video recordings of testimony in legal proceedings.
Under the guidance of professors Happy Mihalca and Mihai Burzo researchers using machine learning techniques, conducted an analysis of 120 videos of actual trials. Based on the final outcome of these legal cases, the researchers initially knew who was involved from the witnesses likely lied during his interrogation.
It was found that among other points to possible false signs, the facial expressions of liars were often changed when they looked right at the person recording their interrogation. In addition, the varied hand gestures, a person’s behavior (for example, the witness started to fidget in his place), increased number of uses of the words-parasites, and awkward pauses, accompanied by “name start” and other phrases, reflecting the potential of the falsity of certain testimony.
By running the information gathered through the written algorithms, the software analyzed the personalities shown in the video. Fifty per cent of witnesses in these videos was telling the truth. Other fifty — lied. Machine could with 75% accuracy to determine which of the witnesses is lying and who is not. Indicator, it should be noted, is very high. Experiments with humans show that people can detect lies only at 50% probability.
The researchers believe that further development of such software by such methods will give different legal structure truly effective tool for detecting lies. Unlike conventional polygraphs, this system does not require physical contact with the investigated person that, in turn, will significantly expand the possibilities of its use.