Why a Grounding in Deception Science Matters

A frequent assumption and misunderstanding of those that hear about Deception And Truth Analysis (D.A.T.A.) and our ability to identify deception and truth in documents and transcripts of the spoken word, is that our algorithm is based on machine learning. It is not, and the reason why is that machine learning – of which we are fans – is statistical, but not scientific. Machine learning is good at identifying correlations that may have been difficult to discover by other means, but it is usually poor at identifying causation, because machine learning is not grounded in science. Instead, at D.A.T.A., our work is grounded in the more than 100 years of deception science and the more than 8,000 papers authored in the space. Here is why a grounding in deception science matters in properly discriminating between deceptiveness and truthfulness…

One way of developing a deception and truth detection algorithm is to use machine learning. Here a vast corpus of texts is gathered together and then a machine learning algorithm is unleashed on the data to discover statistically significant correlations between words or tokens contained in the corpus of texts and future outcomes of interest. These outcomes can include known instances of fraud or deception, as well as things like subsequent: stock price movements, credit ratings changes, earnings revisions, and so on.

Principally, the advantage of this approach is that specific words or tokens can be directly tied to future outcomes statistically. This direct connection keeps the correlation and statistical significance clean. However, these models are not principally concerned with directly measuring deceptive or truthful behavior. 

A disadvantage though, is that if these machine learning identified words do not appear in a text or transcript then the predictability of the signal is nil. For example, if the model is looking for statements about “revenues,” and someone says “sales,” and the model has not accommodated this simile, then the model fails to detect a significant word.

Another disadvantage is that if people discover the words for which the model looks they may alter their communications in order to avoid uttering the “red flag” words. Researchers have, in fact, discovered that red flag words, once identified, become endangered species in future corporate comms.

Nor is the machine language approach principally concerned with or capable of assessing causation. To determine causation you need psychological hypotheses grounded in a scientific discipline such as psychology or criminology where behavior has been studied and forms the basis for scientific inquiry.

These disadvantages are significant, but the biggest practical problem of these models is that they are typically guilty of overfit. They work only as long as future outcomes the model seeks to predict are similar to the training data. But as we all know, future outcomes, especially those tied to stock price movements, are very difficult to predict.

For those that elect to develop deception and truth detection algorithms using machine learning they must constantly fine tune or retrain their models to be assured that they continue to work as future outcomes inevitably change. How frequently to do this is a tricky issue and it does not remove the “we have to predict the future problem” that all algorithms face.

Another drawback is that they likely find themselves in a constant arms race with those who want to avoid using the model’s identified red flag words or tokens. Finally, machine learning-based models must constrain the application to a narrow context, such as news articles, spoken word transcripts, and the like to ensure that there is a close correlation between the trained data set and the future application of the algorithm.

But what if there was a science of deception and truth where the underlying behaviors were better defined?

D.A.T.A.’s algorithm is grounded in the 100+ year history of deception science and its more than 8,000 papers.[i] We then use Natural Language Processing to look for more than 30 known behavioral differences between deceivers and truth tellers that lend themselves well to an NLP approach. In almost all instances the behavioral differences have been replicated in multiple studies.

But why, specifically, is this important? It is important because the behaviors and habits of people once established are very difficult to change. Said another way, a person’s behaviors tend to determine a person’s choices, and these choices lead to measurable outcomes. Thus, if you measure the deceptive and truthful behavior of people then you have the foundations of a more stable and general detection algorithm. A great example is the DATA algorithm’s demonstration of predicting future stock prices despite it never having been exposed to stock price data. Financial markets do eventually price what the executives of businesses, sell-side analysts, and other experts say and do. Here’s why measuring behaviors as identified by science is important when comparing the DATA algorithm to those created using machine learning.

Here is a model proposed by neuroscientists[ii] to describe how decision-making is handled by the brain. Importantly, scientists have demonstrated that because of the historical difficulty of attaining food/energy by hunting and gathering, the brain – an energy glutton – seeks to conserve it. This is the source of habits and behavioral biases, as we explain below. In any case, here is the decision-making model and its three stages:

1. (P)erceptive: information enters the brain through either the senses or metacognition (i.e., self-awareness).
2. (C)entral processing: the brain sorts the stimuli from the Perceptive stage and if they are familiar then it invokes a habit, which may be thought of as an autonomic program. If the stimuli are unfamiliar it triggers more deliberation and much more energy to assess because the regions of the brain that are energy gluttons – like the prefrontal cortex – are invoked.
3. (M)otor: a decision is made, and a course of action taken is evaluated and depending on the outcome it then triggers a physical, hormonal response that reinforces or dissuades the same course of action being taken again in the future depending on the quality of the outcome.

Other researchers refer to this sequence as the “cortico-basal ganglia loop.”[iii] While still other researchers refer to this same brain functionality as a “habit loop.”[iv] Only the nomenclature is different with these different versions of the model, with the three components in this alternative version being: a cue, a routine, and a hormonal reward/punishment.

Once executed, the outcome of the above PCM model is the creation of a cause-and-effect chain that directly associates stimuli (the causes) with actions (the effects). When the outcome of a decision is positive for us then this is reinforced with hormones that really, really make us feel good. This, in turn, biases us to execute this routine the next time, too. If this routine is done frequently enough, it then becomes a habit. And just like a drug, it feels good to engage in the habit. Conversely, negative outcomes of a decision release hormones that lead to negative emotions such as sadness and depression. In turn, these negative feelings lead to us developing a habit of avoidance when future stimuli are similar.

Another reason habits and behaviors are difficult to change is that once they are formed the brain regions needed to execute an action are reduced and consequently the amount of energy needed is also significantly reduced. In other words, the habit-forming hormone-release mechanism exists because the brain is looking to conserve energy. Duhigg says in his The Power of Habit, “When a habit emerges, the brain stops fully participating in decision making.”

Now imagine you are an executive at a business and your company is likely to miss its forthcoming quarterly earnings target. You have several courses of action:

a) Choice 1: You could report the truth publicly. But if your life experience has delivered negative outcomes for you when you have been forthright, then your brain has been routinely flooded with negative hormones that have taught you to avoid these feelings. Additionally, if you are financially incentivized to deliver outstanding results then the feelings associated with financial loss are also likely at play. It takes a strong-willed and exceptional person to overcome these emotions.

b) Choice 2: You could put “spin on the ball” and so that your mostly truthful narrative also has a bit of deception that increases the probability that the earnings miss may be received more favorably. Many of us have this kind of habit and as represented by homilies like: “look on the Brightside,” “there is a silver lining in every cloud,” and so on.[v]

c) Choice 3: You could commit fraud. Here there is no truth to the explanation for an earnings miss, just an alternative and untrue narrative. A study of frauds shows that this is initially very difficult for executives to do as they weigh the consequences of being caught in a lie vs. the emotional pain and possible financial loss of an earnings miss or other negative outcomes. But if the lie is undetected and the overall outcomes are positive, then pleasurable hormones are released in the brain. This makes engaging in deception or lying the next time a little bit easier. In rare instances as we all sadly know; fraudulent behavior can become a habit among company executives.

It should come as no surprise that to change habits is exceptionally difficult due to conservation and a lack of self-awareness. As we learned with the PCM model, above, in the (C)entral processing component, if something is familiar our brains invoke default habits. So, if something isn’t broken, we have no hormonal indication that it needs fixing. In fact, a groundbreaking study sought to establish how hard it is to change a habit and found that the median time it takes to change/create an entrenched habit is 66 days, but with a wide range of 18 to 254 days among participants.[vi] 

Because habits are so expensive to change it means that peoples’ deceptive or truthful behavior is likely already a habit, and thus a reliable predictor. Their entrenched behaviors, in turn, influence to a huge degree their future choices. And ultimately these choices are what drive measurable outcomes, such as the financial performance of businesses. Thus, if you want to predict future outcomes with a high probability you need to measure the entrenched habits of people.

Additionally, we would expect D.A.T.A. Scores which use deception science to assess peoples’ behaviors around deception and truth to be very stable and not volatile in magnitude. In fact, that is exactly what we and independent researchers have found when looking at our history of DATA Scores.

But the truly important point is that because people have a hard time changing their habits a deception and truth detection algorithm based on the findings of deception science has the possibility of being a general deception and truth algorithm. Furthermore, this algorithm is unlikely guilty of the overfit problems of machine learning developed algorithms. This is because our algorithm measures deceptive or truthful behaviors as established by science. In keeping with this article’s financial examples, because markets eventually price deceptive and truthful behaviors the D.A.T.A. algorithm indirectly increases the probability of properly predicting stock price movements. And that is why you should be a DATA Client ????

[i]Denault, Vincent, Victoria Talwar, Pierrich Plusquellec, & Vincent Lariviere. “On Deception and lying: An overview of over 100 years of social science research.” Applied Cognitive Psychology (2022): 1-15

[ii]Vartanian, Oshin and David R. Mandel (Eds.). (2011). Neuroscience of Decision Making. New York and Hove: Psychology Press.

[iii]Graybiel, Ann M. “Habits, Rituals, and the Evaluative Brain.” The Annual Review of Neuroscience. 2008

[iv]Duhigg, Charles. The Power of Habit: Why We Do What We Do and How to Change It. Random House. 2012

[v]At D.A.T.A. we consider there to be important differences between deceptions, lies, and fraud as discussed in our article: “Are Deception and Lying Difference?”

[vi]Lally, Phillippa, Cornelia H.M. Van Jaarsveld, Henry W.W. Potts, and Jane Wardle. “How are habits formed: Modeling habit formation in the real world.” European Journal of Psychology. 40, 998-1009 (2010).