How does one train artificial intelligence to recognize lies and how are neural networks that understand deception different from a lie detector?
1 Does a neural network know what deception is?
About the expert: Andrey Kirsanov, CTO at BrainyS.
First, it is important to mention that “deception” is a very broad concept. If by deception we mean, for example, a fraudulent scheme using some data, then an algorithm can be trained to search for certain sequences of actions and identify fraudsters.
On the other hand, if we are talking about whether a neural network can recognize lies in human behavior, then it is a different and an entirely more complex question.
By itself, a neural network does not know what a lie is. Neural network algorithms can find patterns that are not obvious to humans in the data used in their training. Subsequently, they can use these patterns during the analysis of new data. Therefore, the accuracy of data analysis by a neural network depends on both its architecture and the quality of the data with which it was trained. This also applies to the definition of lies: how well a neural network will detect deception depends is dependent on how the training data containing lies was selected. However, due to the broad nature of a lie itself, it is not so easy to collect data that, in any situation, would unequivocally indicate deception.
2 What data indicates a lie?
Modern science does not have methods by which it’s possible to ascertain a lie as such. We can recognize a lie indirectly through physiological signs, since in some cases deception even affects our body.
There are many variations of lying - from the well-known rescue lie to a professional bluff common among poker players. Everyone, even those with strong principles and beliefs, has used lies at least once in their lives - intentionally or unconsciously. For this reason, different lies have different meanings for people: in some situations, we are deceiving without even realizing it, and in some we make great efforts to make deception go unnoticed.
Accordingly, the amount of cognitive and physiological effort aimed at hiding lies depends both on the level of cognitive load (i.e. how important it is for a person to hide information and deceive the interlocutor at the moment) and on the individual reaction of the body. Depending on how significant a lie is for a person in a given situation, he or she has to make appropriate efforts to conceal the truth and prepare a false answer, and the corresponding physiological parameters change greatly.
By the way, when people who know us well say that we are lying, they usually rely on just these physiological parameters. But even they can be wrong: according to research in the field of deception, people recognize lies with a 50% probability, that is in a completely random way. You might as well just flip a coin.
3 But what about a lie detector? Isn't it more accurate than the human eye?
We often hear about polygraph tests, which are utilized in forensic science. Some companies also use polygraph during job interviews for employees in certain positions. However, the polygraph is just a tool to read physiological indicators. The main part in this procedure is a well-trained polygraph specialist who analyzes the readings of the device and correlates them with their knowledge and experience. Despite the fact that polygraph examiners have years of experience behind them, they themselves assess the accuracy of their work at approximately 70% (more specifically, within a range of 60 to 80%). This means that the external signs and physiological parameters observed during the deception recorded by the polygraph are of a different nature and do not always appear.
Cognitive stress and physical stress can be indicated by changes in various parameters:
- Breathing, usually chest and abdominal
- Cardiovascular activity - pulse and pressure
- Galvanic skin reaction, or skin resistance
- Muscle or physical activity
- Sometimes vocal activity
In studies of reactions during episodes of lying, scientists also frequently record eye movements (saccades), blinking, brain activity, and other indicators.
Why do all these parameters change during deception? Our body reacts in a similar way to any stress. In this case, the nervous system considers the significance of the situation and the concealment of information that may lead to undesirable consequences of stress. It is important to understand that the body's response to a lie is similar to other types of cognitive stress or the manifestation of strong emotions.
4 How can a neural network learn to detect deception based on these parameters? Will it be more accurate than a polygraph?
If a person wants to hide that they are lying, then he or she can try to consciously control their physiological reactions. However, according to research in the field of lies, it is rather difficult to control all physiological parameters simultaneously and consistently. Therefore, modern scientific methods of creating a lie detector are based on the analysis of the entire set of unconscious physiological reactions, including facial expressions. This approach provides a more complete understanding of the state that a person is in, and that is precisely the approach leveraged by the developers of lie detectors based on neural networks.
If we talk about how exactly to teach a neural network to determine a lie, there are two key approaches:
- Give it a video of people lying or telling the truth and mark up this data
- Highlight signs in advance, such as physiological parameters and facial expressions. Based on these signs and the mark up of the video (i.e. at what point a person is telling the truth, and when are they lying), the neural network can learn to recognize when a person is deliberately trying to deceive
It is important to realize that, like a polygraph, a neural network itself does not understand what is true or false, and cannot independently determine that at some point a person has lied. The neural network only shows that the physiological parameters of a person change in comparison with their personal "norm". Already on the basis of these data points, a human operator using a neural network understands that at some moments the cognitive load of a person differs from their “norm” and further interpretation of these points is necessary to for the final opinion of whether the subject is in the process of lying.
Therefore, it is important to take into account that cognitive load and physiological indicators may differ, both because a person is really lying, and because the conversation evokes a strong response from for some other reason (for example, there was a traumatic episode in their life related to the conversation topic).
In summary, the AI system to detect lies becomes an useful tool for an investigator to flag potential points of concern during a lie detection test and combine the results output by the neural network with a degree of personal judgment to make a final assessment of whether deception is taking place.