There's a challenge when you're developing new lie-detection software – you can get people to lie for you in a lab setting, but their behaviour won't be the same as it would be in a real-world scenario. In order to see authentic lying behaviour, you need to go somewhere where the stakes for the liars are higher. That's why scientists from the University of Michigan turned to videos of courtroom testimonies.
Led by professors Rada Mihalcea and Mihai Burzo, the researchers used machine-learning techniques to analyze 120 video clips of actual trials. Based on the outcome of those trials, they knew which witnesses were (presumably) lying when they were on the stand.
It was determined that among other things, liars were more likely to scowl/grimace, look directly at their questioner, gesture with both hands, use filler words such as "um," and use phrases that reflect certainty on their part.
Running algorithms based on that data, the software subsequently analyzed videos of people, half of whom were lying and half of whom were telling the truth. It was able to identify the liars with 75 percent accuracy, which was considerably better than the 50 percent managed by human test subjects.
It is hoped that once developed further, the software could become an effective lie-detection tool in fields such as law, security and mental health. Unlike conventional polygraphs, it wouldn't need to physically touch the people it was assessing, allowing for a wider range of uses.
Source: University of Michigan