If the movie 2001: A Space Odyssey taught us anything, it’s
that computers know when we’re telling a lie. While that may not
actually be the case for most computers in real life, it could
be if they’re running a program created by scientists from the
University at Buffalo. Building on a previous psychological study, the
team produced software that allowed a computer to assess a speaker’s eye
movements, to determine whether or not they were telling the truth in a
prerecorded conversation. It turns out that the computer was able to
correctly able to spot their lies with 82.5% accuracy. According to the
researchers, a trained human interrogator only manages a success rate of
about 65%.


The project utilized 40 videotaped conversations culled from a pool
of 132 used in the original psychological study, in which subjects chose
whether or not to steal a cheque, and were then asked if they had done
so. The selected videos represented a variety of skin colors, head poses
and lighting conditions, plus some of them included potential visual
obstructions, such as eyeglasses worn by the speakers.


In order for the computer to identify the tell-tale “lying eyes” of
each person, it first needed a baseline example of their regular eye
movements, as exhibited when they were telling the truth. This was
accomplished by starting each interview with simple questions, that had
obvious truthful answers, and observing the speakers’ eyes. In
particular, the program took note of their rate of blinking, and the
frequency at which they shifted their gaze.


As the conversation moved on, the subjects were asked whether or not
they had stolen the cheque. If their eye movement pattern remained the
same, it was assumed they were telling the truth. If it changed,
however, they were labelled as liars. While most of the speakers were
caught out, a few were particularly good liars, and were able to keep
their physiological responses (including their eye movements) under
control.


The scientists now plan on further studies utilizing a larger
database, and ultimately hope to develop a system that could be used
alongside human interrogators.


“What we wanted to understand was whether there are signal changes
emitted by people when they are lying, and can machines detect them?”
stated assistant professor Ifeoma Nwogu. “The answer was yes, and yes.”