A Machine Can Read Lying Faces Better Than Humans
Human beings are not great at lie detection. One study found that people can guess when someone is lying just a little more than 50 percent of the time. However, machines may be able to catch us lying more often. A group of researchers working at the University of Michigan have developed an algorithm that can detect lies from videos at higher rates than human interrogators.
The algorithm, developed by Rada Mihalcea, a professor of electrical engineering and computer science at the University of Michigan, can detect lies around three quarters of the time. The algorithm was trained on real videos from courtrooms, including statements from exonerated prisoners from the Innocence Project who had been found to be wrongfully convicted of crimes, and television shows like Lie Witness, which show interviewees who are known to be lying (for instance, inventing an opinion about a movie that doesn’t exist). The algorithm analyzed verbal and nonverbal behavior, including gaze, facial expressions, hand movements, the complexity of the person’s syntax, and more.
In one experiment, Mihalcea’s lie detector outperformed people by up to 15 percent, identifying lying faces with as much as 77 to 82 percent accuracy. But the study did have some limitations: In the courtroom videos, interviewees were considered to be lying or telling the truth based on the verdict of the trial, which obviously introduces some room for error, since juries are not always correct (hence the existence of the Innocence Project, which investigates wrongful convictions).
Mihalcea and her colleages’ work was presented at the Conference on Empirical Methods in Natural Language Processing in September [PDF], and will be presented at the International Conference on Multimodal Interaction later this month.
[h/t: New Scientist]