We think we may have discovered a better polygraph.
Telling truth from lies is an ancient problem; some psychologists believe that it helped drive the evolution of intelligence, as hominids who were better at cheating, or detecting cheating by others, left more offspring. Yet despite thousands of years of practice, most people are pretty bad at lie detection, and can tell lies from truth only about 55% of the time – not much better than random.
Since the 1920s, law enforcement and intelligence agencies have used the polygraph, which measures the physiological stresses that result from anxiety. This is slightly better, but not much; a skilled examiner may be able to tell truth from lies 60% of the time. However it is easy for an examiner who has a preconceived view of the suspect’s innocence or guilt to use a polygraph as a prop to help find supporting “evidence” by intimidating them. Other technologies, from EEG to fMRI, have been tried, and the best that can be said is that it’s a complicated subject. The last resort of the desperate or incompetent is torture, where the interviewee will tell the interviewer whatever he wants to hear in order to stop the pain. The recent Feinstein committee inquiry into the use of torture by the CIA found that it was not just a stain on America’s values but ineffective.
Sophie van der Zee decided to see if datamining people’s body movements might help. She put 90 pairs of volunteers in motion capture suits and got them to interview each other; half the interviewees were told to lie. Her first analysis of the data was to see whether you could detect deception from mimicry (you can, but it’s not much better than the conventional polygraph) and to debug the technology.
After she joined us in Cambridge we had another look at the data, and tried analysing it using a number of techniques, some suggested by Ronald Poppe. We found that total body motion was a reliable indicator of guilt, and works about 75% of the time. Put simply, guilty people fidget more; and this turns out to be fairly independent of cultural background, cognitive load and anxiety – the factors that confound most other deception detection technologies. We believe we can improve that to over 80% by analysing individual limb data, and also using effective questioning techniques (as our method detects truth slightly more dependably than lies).
Our paper is appearing at HICSS, the traditional venue for detection-deception technology. Our task for 2015 will be to redevelop this for low-cost commodity hardware and test it in a variety of environments. Of course, a guilty man can always just freeze, but that will rather give the game away; we suspect it might be quite hard to fidget deliberately at exactly the same level as you do when you’re not feeling guilty. (See also press coverage.)
10 thoughts on “To freeze or not to freeze”
Not sure I accept your closing observation. It would no doubt require a degree aptitude and training, but I am pretty certain the bad man or woman, depending on what message they wished to convey, could freeze or fidget in just the right way.
Of course gamblers have known for millennia that body movements can be read. Good poker players often manage to fool other players who are themselves highly experienced at reading behaviour. Interestingly, even the best players can eventually be read by equally skilled players.
Thanks for your responses. We are currently designing the next set of validation studies, in which we will deal with several issues pointed out in your comments. Thanks for your input, really appreciate a critical point of view. One of these studies will be with convicted sex offenders, giving us the opportunity to test how well this method works on people with different personality traits. That will be very interesting.
Also, I fully agree with the point that once someone knows that people move more when lying, that person is likely to adapt his behaviour to try to get away with lying (in this case by limiting his movements, or to “freeze”). There are 2 reasons why I personally think that moving the exact right amount (as an indication of truth telling) might be hard. First, we found this movement difference between truth tellers and liars in all separate limbs we tested (e.g., arms, head, trunk and legs). And although you can purposefully limit your hand movement when paying attention to it, it might be harder to for example limit trunk movement because you’re less aware of those movements. Or least, restricting movements in all body parts at the same time as speaking might be challenging. Second, we’re currently working on creating software that analyses movement in real-time (i.e., whilst it’s happening), and we’re considering to create a “green/truth signal”, a “red/lie signal” and an “amber/suspicious” when someone moves abnormally little or way too much (which might be a sign that he’s trying to beat the system). Of course, with an accuracy rate of 82.2% I don’t suggest to use this system as evidence in court – it’s not good enough for that – but it might be useful as an aid during interviews, just an indication that you might want to ask this person a few more questions on this topic. However, to only way to know for sure what effect “countermeasures” (trying to beat the system) will have, and if we can detect them, is by testing this, which is what we’ll do in a systematic manner. Once finished, we’ll of course report back here on LBT.
I am tempted to agree with Keith above, although my reasoning is quite subjective – there are after school clubs for 5 year old kids that do – drama, communication skills, dancing, etc. All these skills teach children how to control their body. There are also many courses around for client-facing people that teach the same over a few weekends.
It is definitely a good piece of research, showing that you were able to find a classification method that achieves a good result for a particular set of 90 samples, and the paper seems to demonstrate that a lot of effort has been put into finding this method. I would hesitate to say much more about results where a success is argued without use of a control group, A/B split, replication … although I’m definitely not an expert on true nor quasi-experiments.
I wonder what is the level of correlation here, i.e., do we learn more after 10 interviews with a given person than we knew after the first one?
How did this get published at Cambridge? what a joke.
All this paper amounts to is: there’s a weak correlation between fidgeting and deceit. Not a ‘reliable’ one. It’s interesting information but not enough to be indictable—that needs to be emphasized.
Maybe we should add the clarifier ‘in normal people’. Try it on the stone cold sociopaths who brutally murder people. They don’t feel guilt over what they do.
The paper was one of two we published following peer review at HICSS, the conference where people send their best results in deception-detection technology. If some anonymous trolls don’t like that, tough; peer review is the quality assurance mechanism we academics use.
There are further press articles in Dutch here, here, here, here, here, here, here, here, here and here.
I find the results intriguing, but I wonder if you took into account the number of people that face criminal charges that also have undiagnosed disorders such ADD/ADHD. Even though the physical hyperactivity tends to fade in adult ADD/ADHD, it’s still present in both women and men.
It is a well-documented phenomena that those with the disorder are more likely to commit crime and serve jail time (the theory being that ADD/ADHD is partially an under-development or malfunctioning impulse control system) and that a potentially large minority of current prisoners have ADD/ADHD.
So if a large minority of people that commit crime suffer from a disorder that is rarely screened for in most adult populations and is highly correlated with poor impulse control, will the device be able to create a baseline and then test for guilty fidgeting from there?
Interesting research, but still, assuming you do “improve” your technique to working 80% of the time, this will be of small comfort to 1 in 5 of those falsely suspected or accused and wrongfully condemned by its application.
This is interesting in principle, but I would have expected something as predictable as freezing/fidgeting to be recognised reliably by people – we are, after all, good pattern-recognition engines.
I am interested in Lauren’s comments above, but from a slightly different set of experiences. Psychotropic medications for serious metal illness (at least the older phenothiazines) tend to affect body movement, and it can be difficult to “read” the veracity of people on these medications.
Now there is a short video as well.
Here at last is the official journal version of the paper.