A newly developed lie detection program fares better than most humans, its inventors have recently announced.
Lately, machine learning has proven successful in analyzing big data in order to issue reliable predictions, and it has also played an instrumental part in creating autonomous vehicles and voice recognition software.
The process basically consists in allowing computers to build new skills and boost their artificial intelligence by studying information, in order to identify patterns and future courses of action without having to be programmed by their users. Now it appears that another application has been found for machine learning: lie detection.
Experts at the University of Michigan created software so as to serve this purpose and perfected it using massive amounts of courtroom data, so as to improve its accuracy.
Some of the information was provided thanks to representatives of the Innocence Project, a non-profit organization whose purpose is to vindicate wrongly convicted individuals, through DNA profiling.
By meticulously analyzing 120 videos showing court proceedings, it was possible to identify clear tell-tale signs that can indicate a person’s guilt or innocence, by determining whether someone is spurting lies or being truthful.
For instance, some of the key indicators that the software came to rely on were patterns that can’t be normally tracked by ordinary people, such as the frequency of certain gestures, the number of verbal fillers and pauses, the amount of eye contact, or the intonation used when expressing various thoughts or recollections.
More precisely, deceitful persons are much more likely to fidget with their hands, to stress words in order to make them sound believable and to hold someone’s gaze for excessive amounts of time in an attempt to dispel doubt.
In addition, it’s also more probable that they will stare angrily and grimace, use fillers such as “um” or “uh”, and attempt to separate themselves from the events by recounting at the 3rd person instead of the 1st.
Researchers then conducted an experiment in order to test the software’s lie detection abilities against that those pertaining to humans, by comparing guesses regarding a defendant’s guilt with the actual verdict.
Eventually, the conclusion was that the program’s effectiveness when making such inferences had been of approximately 75%. In contrast, it was also discovered that humans were much less adept at spotting liars, their guesses being faulty around half the time.
This is hardly surprising for Rada Mihalcea, professor of computer science and engineering at the University of Michigan. As she explained, people focus on the bigger picture when assessing a person’s truthfulness, and aren’t capable of actually counting the number of certain nonverbal signals, so as to detect potential incongruities.
Computers however have this ability, which makes them more proficient at detecting lies in comparison with their counterparts.
Researchers are now planning on developing an even more reliable software, which will integrate elements of a traditional polygraph, by tracking the subject’s pulse, respiration, blood pressure and electrodermal activity (since excessive sweating can result from psychological turmoil).
Ideally, such indicators should be monitored without the participant being aware of having been put under the microscope, using thermal imaging or other similar techniques.
Moreover, another objective is to give computers greater freedom in assessing if a certain motion is suggestive of guilt or innocence.
It may be that humans’ understanding of gestures signalling dishonesty is actually flawed and incomplete, so it would be recommended to test even long-held beliefs about the body language of a liar.
Eventually, such software could be employed in courtrooms, but also in airports and other vulnerable locations requiring extensive and reliable security checks.
Image Source: Flickr