Exploring the truthfulness of the polygraph

I think most people would agree that without comments, Habr would be a completely different place: less interesting and informative. They often contain important additions and clarifications, they tell interesting stories and, of course, they debate (we are not in parliament, right?). Some debates are more interesting than the articles themselves. And just recently it happened again: I saw a comment that I disagreed with. My reaction to this is usually simple: put a minus in karma and close the tab try to express your position and listen to the opponent's arguments. I did the same in this case, but then it turned out that the question was much deeper than I thought – I would have to choose one of the modifiers “quickly” and “reasonably” to answer. Besides, maybe I'm completely wrong? In short, I wanted to understand the issue.
For those who don't follow the links: the discussion was about lie detectors (also known as polygraphs) and their functionality. In this publication, we will try to figure out what kind of machines these are, how they are used, and whether they can be trusted. For these purposes, we will talk primarily about AI:

Idea and history

The idea is, in essence, simple: telling the truth is easy and pleasant, but lying is not. And when something is unpleasant for a person, he begins to breathe quickly, sweat, his heartbeat quickens, and his eyes start darting around. On an intuitive level, everything seems simple, and if so, then it can be safely used in practice. Surely you have heard stories about criminologists from a thousand years ago (or even more!), like these:

• A person puts dry rice (or rice flour in other versions) in his mouth and listens to the accusation. If the rice remains dry, then the dog is to blame: he began to worry and because of this he stopped salivating.

• A hot blade is placed on a person's tongue. If it sticks, don't expect anything good. The principle of operation is apparently the same as in the previous example.

• A certain number of people pass an egg with a fragile shell back and forth, and the one who is guilty gets nervous and crushes it.

• The suspect is told words in random order, and he stands and quietly beats the gong. But as soon as he hears a word related to the crime, he will hit the gong harder because of nerves: that's when he needs to be tied up.

• We tried measuring the pulse: if it is rapid, then the person is lying.

• They carefully looked at the suspect's face and looked for signs of deception.

And other stories that exude mountain coolness. All of this sounds at least minimally justified, but I was never able to get to the source of even one of these methods. In both Russian and English, it seems that these methods are retold in a circle.

Slightly less obvious ways to detect lies

This is not directly related to the matter, but since I found it anyway, let it be here.

Separately, it is probably worth highlighting ordeal — lie detectors come from the Middle Ages. Suspects were asked to do something simple: pull a ring out of boiling water, hold hot metal in their hand, or walk on hot coals. And then either the person remained unharmed (divine forces, of course, will help the innocent), or one of two things. Trial by combat is the same: would providence allow a liar to be defeated? However, sometimes the degree of cruelty was reduced, we know about the ordeal with bread and cheese: the suspect was tightly girded with some rope around the stomach and had to eat a significant amount (you won’t believe it) of bread and cheese. Sometimes they also forced him to drink water. The most humane, perhaps, was the ordeal by miracle: it was enough to kiss the cross or visit the grave of a saint. All sorts of satamans and drug addicts, of course, could not do this.

Another method that came from India was used: priests would smear the tails of several donkeys with soot and lead them into a dark room. Suspects were asked to enter and pull the donkey's tail. If the donkey brayed, then the guilt was considered confirmed. And if there were no sounds, but the person's hands were actually clean when they left, then this meant that they were metaphorically dirty: he was afraid to pull the tails!

In Rome there is a slab called neither more nor less The Verità Basin. I mean “The Mouth of Truth”. I think it's pretty clear from the picture: you have to stick your hand in there and if everything is good, then everything is good. And if not, then your hand will be bitten off, of course. There's a scene that plays on this in the movie “Roman Holiday”. They say it's good, but I haven't seen it myself, so I can't say for sure.

The Babylonians invented a lot of things: dividing a circle into 360 degrees, and a year into 12 months, a bit of math, watches. Beer, which is important. And also this way to identify a liar: it was believed that he looked at the ground (this is understandable, okay) and drew circles with his big toe (apparently some kind of cultural feature).

It is probably unnecessary to mention any methods of the “let’s cut the ram and figure out what’s what by the guts” variety.

In short, the idea was roughly the same then and now: in a calm state, a person functions in one way, and when lying, in another. And if the differences between these states are described and recorded, then you can understand where you are being lied to. Based on this, Lombroso (not the last person in forensics) suggested, for example, measuring blood pressure. And he even seemed to have exposed someone in a lie.
What problems can be seen here? There are two obvious ones.
First — Did we count the pulse correctly, measure the pressure and all that other stuff? Are our devices accurate and sensitive enough?
Second – And who knows, why did the donkey bray and the man's pulse quicken? What does that mean?
Let's remember them. This was all, however, just a preamble.

What, how and why does a polygraph measure?

There are no uniform standards, but on average the list of sensors is approximately as follows:

• Pneumograph (respiratory movement recorder).

• Pressure gauge.

• Galvanometer for measuring skin conductivity (read – amount of sweat).

• Activity sensors (something like pillows that you put under your arms, legs, and, excuse me, butt). They are also called “countermeasure detectors”, or something like that.

• Sometimes some exotic things like a plethysmograph (this is the first time I’ve heard of such a thing, to be honest).

  I'm not talking about any peripherals like microphones and screens, where all the indicators are displayed.

When assembled, it all looks something like this

The procedure itself, again, on average, looks like this:

• First, they will just chat with the person: they will tell him about the procedure, give him a list of questions that will be asked (by the way, there is no consensus on this: some say that you need to give him all the questions verbatim, while others think that it is enough to say, well, we will ask how to patch kde2 under freebsd about this and that, you don’t need to know more specifically.

• The, uh, subject is dressed in all this gear and is somehow asked some questions for calibration. Why “somehow”? I used to think that they were asked in the usual mode, but during the research I found this, for example: “You'll be asked questions in both an accusatory and non-accusatory manner to record your reactions to both approaches”, that is, something like “You will be asked questions in both an accusatory and non-accusatory manner in order to record your reaction to them”. Why “some”? There is no uniformity here either. Some people think that you need to chat about something neutral, like a hobby, and go through a list of target questions in a free form, while others simply ask the same questions that will be in the control (d)eskment, only, perhaps, in a different order. The only uniformity is that you will be asked something known in advance, for example, a name or a day of the week in order to record the reaction and calibrate it. Oh, and they will also ask whether the person is sick with something that is incompatible with testing: cardiovascular diseases, for example. Or epilepsy.

• After this, the actual testing begins. “After” should be understood as “from several dozen minutes” to “several weeks” depending on the work schedule of specially trained people and the type of test (more on this below). One way or another, the subject is asked questions (interesting and control), several times for the sake of accuracy. The questions are formulated so that they can be interpreted unambiguously and answered “yes” or “no”.

• Finally, a specially trained person looks at the graphs that the equipment has drawn and makes some conclusions. He can ask additional questions if he has any doubts. He can tell about the results, or maybe not, it depends on the customer, as I understand it. That's basically it. The amount of time, by the way, for the active part of the procedure also varies, the range is quite large: from thirty minutes to four hours.

And here we are

to the most interesting: what is the accuracy of all these procedures? Who should I ask?
If you ask polygraph examiners, the accuracy is simply off the charts, we are talking about numbers in the region of 90%. Where such numbers come from can be understood from this meta-review, for example. I'm not a big connoisseur of statistics, but without reading very deeply I did not find any obvious manipulations or omissions. However, it is worth remembering that this study is to a certain extent biased and the criteria for selecting articles for the review could have been adjusted so that what was needed would get in. And what was not needed would not get in. Something in the spirit of the film “Thank You for Smoking”.

If you ask someone else, it's much harder to find numbers. But doubts are much easier. What's the matter?
There are a whole set of reasons that can be identified here. I will try to list them. Remember how we remembered two obvious problems? It's time to remember.

Accuracy of instruments

On the one hand, modern technologies allow us to manufacture fairly accurate sensors that can measure the parameters we consider essential with reasonable accuracy. On the other hand, it is not entirely clear what threshold a change in a parameter is above becomes essential: if we set it too sensitively, we will not find two similar reactions; if we set it too sensitively, we will get a result in which all reactions are essentially the same. To some extent, this problem is solved by calibration and a number of questions, but we cannot say that it is completely eliminated: there is always a chance of making a mistake.

Causality

That same question about why the pulse quickened. It seems obvious that our thoughts can influence how we feel: try to think about work and you'll get a headache Think of lemons and you will most likely start salivating immediately (at least for me). But this does not mean that absolutely all people react to the same thoughts in the same way (people who do not salivate from lemons can leave a comment). And so it is not entirely justified to simply claim that everyone who lies has a faster heart rate or sweats. Again, the polygraph test itself is most often a stressful situation (at least because it is atypical). Preliminary testing is intended to partially eliminate this problem, but this can be compared to rehearsing a speech and the speech itself: no matter how much you rehearse, some anxiety will still remain. In short, two objections were born from the original question: “Why do we think that everyone reacts the same way when telling a lie?” and “How can we separate natural anxiety from anxiety when lying?”

Ratio of responses to control and relevant questions

Control questions, let me remind you, are those whose answers are known in advance: “What color is the sky?” or “Have you ever lied?” And relevant questions are related to the study: “Didn't you chop the old lady with an axe?” And it is assumed that the reaction to control and relevant questions differs enough to understand that you are trying to lie on the relevant one. But it seems pretty obvious that if you are suspected of murdering an old lady, then even if you are innocent, this question will evoke a much more vivid reaction than the question about the color of the sky.

Umbrella term

All this time, speaking about the polygraph, I meant the complex of sensors that I described at the beginning: all these plethysmographs and so on. But generally speaking, a polygraph test can mean a fairly wide range of activities. Just fourteen years ago, on Habr, they talked about portable lie detectors that are limited to measuring skin conductivity.

The elephant in the room

We haven't really touched on this aspect before, but the polygraph itself as a system of sensors doesn't determine anything. The results are interpreted by a person who frowns as he looks at the graphs. And the decision that here there are deviations, but they are normal, and here the person may be hiding something is also made by him. I looked for some data on automatic Satisfiers polygraphs and I didn't find anything in particular in either Russian or English. Only these Guys.
And if there is no possibility to single out some general principle of work, then I shake my head in disbelief. Perhaps, however, I am wrong and this is a lobby of polygraph examiners who simply do not want to remain without work.
Yes, by the way, it is worth saying that in order to be a polygraph operator it is not necessary to have any special medical or psychological education: it is enough to undergo several months of training.

Instead of conclusions

Unfortunately, I can’t say anything definite based on the results. Partly because I see contradictory conclusions from seemingly reasonable premises, partly because most of the studies I found were conducted some bearded years ago, half a century ago, and there aren’t really any modern ones. It seems that the scientific community has either lost interest in the topic or decided that the results obtained at one time are enough to draw definite conclusions. It’s probably worth agreeing that when conducting a (d)survey under ideal conditions with the help of super-professionals, it is entirely possible to achieve some statistically significant indicators. In reality, however, the conditions are not always ideal, and super-professionals are not always the ones working. In addition, the person conducting the study always operates with probabilities and never says “yes, this is definitely the truth” and “no, this is definitely a lie.” So, regarding the general case, my opinion has probably not changed: a polygraph test is about as reliable as a throw of a dice.
Here I did not mention Aldrich Ames (he worked for the CIA, he probably knew secret techniques for cheating the polygraph) and new methods of detecting lies using artificial intelligence. Perhaps, in my free time, I will study them and write something.
In the meantime – subscribe to my telegram channel Vote in the poll and write comments if you (dis)agree with what is written.