AI System DARE Can Detect Lies in Courtrooms
American developers have created an artificial intelligence system called DARE that can detect when a defendant is lying by analyzing their facial expressions and voice. Testing has shown that the program outperforms humans at this task—DARE correctly identified lies in 90 percent of cases. The program will be presented at the AAAI-18 artificial intelligence conference, and a description of its principles can be found in a preprint on ArXiv.org.
Humans Struggle to Detect Lies
In everyday life, people are not very good at recognizing deception. Studies show that, on average, untrained participants can correctly determine whether someone is being honest only 54 percent of the time. However, lies can be revealed by various signs—not just physiological changes like increased heart rate or sweating, but also external characteristics.
For example, even slight changes in facial expressions can reflect emotions a person is trying to hide. Modern machine learning systems have great potential for detecting deception, largely because computers can be trained on large datasets, allowing them to notice subtle details that humans might overlook.
How DARE Works
Researchers from the University of Maryland and Dartmouth College developed the DARE algorithm, which acts as a kind of polygraph. DARE is based on several existing, pre-trained classifiers and models. To train the algorithm, the team used a database containing 121 video clips of simulated court proceedings. The system tracked changes in facial expressions, voice, and speech. The program’s workflow is illustrated in the image below.
Testing and Results
The developers evaluated DARE by showing it videos of actors who were either lying or telling the truth. During cross-validation, the algorithm achieved a maximum accuracy of 92 percent (compared to 81 percent for humans). Interestingly, the main contribution came from external cues, such as furrowed brows or pouting lips. When relying solely on changes in facial expressions (without other context), DARE detected lies in 87 percent of cases.
Previous Research and Limitations
A prototype of a similar program using the same database was created back in 2015 and presented at the International Conference on Multimodal Interaction, which focuses on human-machine interaction. Testing showed that the earlier program achieved 75 percent accuracy. However, the authors of both the new and previous studies point out one limitation—the systems are based on simulated court proceedings. In real life, defendants may be much better at lying, since their motivation is radically different from that of someone participating in an experiment.