The ability to detect with high accuracy if someone is lying would be extremely useful in many scenarios, but most obviously law enforcement. The idea is so alluring that we collectively just cannot give up on it, no matter how elusive the technology has turned out to be. The media typically presents the issue as a matter of lie-detection technology, and likes to hype new fancier technology as if that is going to be the path toward highly effective lie detection. I will explain below why that is likely not true, but first lets take a look at the latest such reporting.

A recent BBC article, True story? Lie detection systems go high-tech, is typical. They highlight a study from researchers at Tel Aviv published in December 2021 that uses electromyography to measure facial muscle movements as a telltale sign of lying. The BBC’s summary of this study is essentially that the researchers found there are two types of liars, those that move their mouth when they lie and those that move their eyebrows. Further, the research found that lie detection based on this method is “73% accurate”.

There are multiple methods for machine-enhanced lie detection. The classic lie detector is based on measuring physiological parameters, like heart rate, breathing and sweating. These are essentially measures of anxiety, and so this type of lie detector is really an anxiety detector, based on the assumption that people who are lying are more likely to be anxious than someone telling the truth. However, there are good liars who do not get anxious, and there are people who get anxious because they are being grilled, whether or not they are guilty of something. So these type of lie detectors have a high false positive and false negative rate.

There are also lie detectors based on fMRI examination. The idea here is to find areas of the brain that become active when we are lying but are not active when we are telling the truth. This seems like the holy grail of lie detectors – you can’t control what your brain is doing at a subconscious level. But of course, this is not really true and there are inherent limits to this technology. Actors, for example, sometimes use a method where they get themselves to actually feel the feelings they are trying to act out, to make them more genuine. Perhaps a good liar can, on some level, convince themselves that whatever narrative they are pushing is reality. Also, some people can be relatively indifferent to the truth, and are not so much lying as saying whatever they want to be true without distinguishing whether or not it is true.

But perhaps the greatest limitation of fMRI lie detectors is the technology itself. When you see those pretty pictures of areas of the brain lighting up during a specific task, those are composite images, of many trials and even many participants. There is a lot of background noise on fMRIs, and it is often challenging to pull a signal out of that noise. This makes them useful for research, where averaging lots of trials can still be meaningful, but hard to apply to an individual, where noise will predominate.

The new study uses a different method entirely – looking at facial muscle movements (microexpressions) to detect lying. Past approaches using this method used a camera to visually inspect and measure facial movements. What’s new about this technique is that it uses EMG – electrodes to measure the contraction of specific muscles. They placed electrodes on one side of the face (assuming that facial movements would be bilateral, so only one side was needed), and focused on the superciliary muscle above the eyebrows and the zygomaticus muscle of the cheek. The results are somewhat interesting, but in my opinion reveal the ultimate limitations of lie detector technology.

They ran trials with subjects who were give one of two phrases and then has to state to a receiver either the correct phrase that they heard or the other phrase they did not hear. They were told to convince the other person that the phrase they tell them is the one they heard, and the receiver was tasked with determining if they were lying or not. In a second round of experiments the subjects were give cash if they were successful, as an attempt to enhance their motivation. Without any aid from a lie detector, the receivers on average were about 50% successful in guessing if the sender was lying – no better than random guessing. This translates to a net zero ability to detect lying in this scenario.

They used this data to train an AI algorithm to cluster facial movements that correlate with lying, and then use that as a predictive model to detect lies. The algorithm ultimately found that when people lie they fall into one of two main clusters, those who raise their eyebrows when they lie and those who move their cheek when they lie. The main question, of course, is – how did this EMG-based lie detector do? The results are unimpressive.

“We found that the classifier was significantly better at detection than humans (i.e., Receivers) both in the first and second stages of the experiment (t(39) = −6.8, p < .001 and t(39) = −8.4, p < .001, respectively)”

Essentially they rated how successful senders were at lying to people and then compared that to how successful they were against the lie detector. They were 6.8% less successful in the first trial, and 8.4% less successful in the second. This is a somewhat odd way of presenting the data, and that always makes me suspicious that some P-hacking was going on. Were other ways of parsing the data not significant? What I was initially looking for was a straight-up measure of the false negative and false positive rate, and I always find it disappointing when I can’t find that most useful of measures. Often the data is presented as how “accurate” the predictions were, but that is not good enough as it is ultimately ambiguous.

The problem, however, may not be in how the researcher chose to analyze and present the data, which may not represent any deception or P-hacking on their part. The odd way of presenting the data may ultimately be due to the inherent underlying problem of this entire endeavor – there is no one consistent behavior that is a marker for lying.

The researchers in this study acknowledge as much in the discussion:

Our individual-level, within-participant analyses were designed to identify key factors that contribute to detecting lies of specific individuals—both temporal events (when people heard the stimulus, delivered the message, or awaited a response) and facial muscles (ZM and CS). The fact that we identified different types of liars goes against the idea that expression of deceit has universal indicators

Even more significant – individual people changed their “tell” in later trials. People were different from each other, and from their earlier selves. This all sounds like chaos, and elaborate methods of teasing out some statistical signal from this noise calls into serious question if there is a real phenomenon here, and can it be applied to individuals in the real world.

I also have a problem with the research paradigm here, as it does not duplicate the stress of real-life lying when there are high stakes. This gets back to the anxiety problem. Peoples tells might also vary in different situations, under different stress levels, and perhaps when talking to different people. Data from lie detectors may be noisy because the behaviors surrounding lying (which itself is a spectrum) is inherently noisy. Better technology will not solve this problem, just provide different methods of detecting the noise.

Which means that as a practical device, lie detectors may be stuck with relatively low sensitivity and specificity, which severely limits their utility in legal matters and calls into question the ethics of relying upon them for important decisions. What I have been repeatedly told by law enforcement individuals is that it can be a useful tool in interrogation, essentially as an intimidation tactic. If people believe their lies can be detected, they may crack. But that use does not even rely upon lie detectors working, only a popular belief that they do.

This latest study does not really move the needle much on the technology of lie detecting, and for me reaffirms the basic limitation of this approach. The behavior of lying is too inherently variable and complicated to apply any binary physiological or behavioral measure for its reliable detection.