Polygraph tests, more commonly known as lie detector tests, make for great drama on your favorite crime show, but the reality of their usefulness is far less believable. Already rejected by the court system as admissible evidence due to their lack of reliability and their overall failure rate, polygraphs are still used for other tasks, like candidate screening and narrowing a pool of suspects. One police detective has stated that the test’s only useful criminal function is in helping rule someone out just by his willingness to take the test: if he’s not guilty, he’ll jump right in the chair, but if he has something to hide, he’ll refuse to comply.

fingerprint

 

Basically, there are too many factors involved in screening a suspect–as well as too many variations in how people respond to the questions–to make physiological response-based testing accurate. But researchers at the University of Michigan are taking a different approach, and producing a different type of lie detector test based on machine learning software. One of the chief differences is that the software never touches the suspect, but bases its prediction on other behavioral features.

Programmers examined hours and hours of video clips of actual trials, then transcribed the audio of those clips. They also noted every facial movement, hand movement, and quirky tick of the witnesses on the stand. They then compared those to the trial outcomes, which they used as the measure of truth or lie, and “taught” the computer to correlate the behaviors and speech patterns of the witnesses to the “truth,” or in this case, the verdict.

Of course, there are inherent flaws in this methodology, namely that a verdict isn’t (unfortunately) a completely solid indicator of the truth. Justice advocates have long criticized the bias in our legal system and the lack of total transparency when it comes to determining guilt. But this is a small step in the right direction: while traditional polygraph testing produces a pitiful 50% accuracy rate (essentially the same as throwing a dart to pick a suspect, considering many tests rely on a 50-50 yes/no question system), the software that researchers are developing already has a 75% accuracy rate, which can be anticipated to improve with more sample studies to learn from.

One of the most telling signs in favor of this software is the cooperation of The Innocence Project, an organization that works to undo the faulty verdicts of inmates whose cases have significant errors in the way the investigation and trial proceeded. The organization supplied the researchers with some of the video clips used to train the software in what’s true and what’s not.