Detecting deception with artificial intelligence: promises and perils

Rapid advancements in artificial intelligence (AI) have driven interest in its potential application for lie detection. Unfortunately, the current approaches have primarily focused on technical aspects at the expense of a solid methodological and theoretical foundation. We discuss the implications thereof and offer recommendations for the development and regulation of AI-based deception detection.

Orienting versus inhibition: The theory behind the ocular-based Concealed Information Test

When trying to conceal one's knowledge, various ocular changes occur. However, which cognitive mechanisms drive these changes? Do orienting or inhibition-two processes previously associated with autonomic changes-play a role? To answer this question, we used a Concealed Information Test (CIT) in which participants were either motivated to conceal (orienting + inhibition) or reveal (orienting only) their knowledge. While pupil size increased in both motivational conditions, the fixation and blink CIT effects were confined to the conceal condition. These results were mirrored in autonomic changes, with skin conductance increasing in both conditions while heart rate decreased solely under motivation to conceal. Thus, different cognitive mechanisms seem to drive ocular responses. Pupil size appears to be linked to the orienting of attention (akin to skin conductance changes), while fixations and blinks rather seem to reflect arousal inhibition (comparable to heart rate changes). This knowledge strengthens CIT theory and illuminates the relationship between ocular and autonomic activity.

No effect of moderate alcohol intake on the detection of concealed identity information in the laboratory

The Concealed Information Test (CIT) enables the detection of certain (e.g., crime-relevant or personal) information, even if participants aim to conceal their knowledge. The current preregistered study investigated whether previously observed impairing effects of alcohol intoxication on participants’ performance in a reaction time CIT (RT CIT) field study also translate to a laboratory environment. In contrast to the previous study of Suchotzki and Gamer (Sci Rep 8:7825, 2018) in which alcohol consumption was voluntary and self-administered, the current study used a randomized assignment of participants to either an alcohol group (n = 88; receiving a drink with 3 cl alcohol) or a sober control group (n = 89; receiving a drink with just some alcohol drops to hide group assignment). After drink administration, participants completed an RT CIT, in which they were instructed to hide knowledge of their own identity. Blood alcohol concentration (BAC) was estimated via breath alcohol ratio. In contrast to the previous field study, results revealed no differences in CIT-performance between intoxicated and sober participants. Aside from questioning the robustness of the result of the previous field study, our results also point to a number of interesting theoretical explanations for the discrepancy between both results, which are elaborated in the discussion.

Testing the applied potential of the Sheffield Lie Test

The Sheffield Lie Test (SLT) has been frequently used in laboratory research investigating basic mechanisms of deception. Its applied potential as a lie detection tool has been contested. The current two experiments used a reaction time SLT and investigated whether it can discriminate between participants who committed a mock crime and participants who performed an everyday activity. Results of the first experiment revealed that guilty participants (n = 32) took longer and committed more errors when having to deceptively deny the mock crime and deceptively confirm having performed the everyday activity in contrast to truthfully admitting the mock crime and denying the everyday activity. Innocent participants (n = 29) showed the reversed pattern. Individual Cohen's d's and the area under the ROC curve revealed a high above chance discrimination between both groups. In a second experiment, we repeated this procedure, yet participants were now given a more elaborate explanation of the alibi activity that all should pretend to have done. Although results still revealed the expected pattern in innocent participants (n = 48), the effect was not significant any more for the guilty participants (n = 46). Accordingly, classification accuracies also dropped. These two experiments demonstrate the applied potential of the SLT, yet at the same time its severe limitations. Potential solutions and suggestions for future research will be discussed.

Effect of negative motivation on the behavioral and autonomic correlates of deception

In forensic contexts, lying is often motivated by the will to avoid negative consequences (e.g., an arrest). Previous research investigating the effect of motivation on deception has, however, nearly exclusively focused on the effect of positive motivation (e.g., via financial rewards). In the current study, we aimed to replicate previous studies on the behavioral and autonomic correlates of deception and to investigate the influence of negative motivation on those correlates. Participants committed a mock theft and underwent a questioning procedure in which they had to truthfully and deceptively answer questions about the committed mock crime and a control crime. Half of the participants completed the procedure without any specific motivation, whereas the other half of the participants was told that lies detected by a computer algorithm would result in an unpleasant electric stimulation. Results revealed longer reaction times, larger skin conductance responses, and a relative heart rate slowing for lying compared to truth telling. The latter two effects were larger in the motivation compared to the control condition. Results are in line with the motivation impairment hypothesis and effects of positive motivation on the autonomic correlates of information concealment. The observation that a higher motivation increased rather than decreased deception effects is promising for forensic contexts. The possible explanation that these effects could be related to fear conditioning should stimulate future research on the stability of differences between lying and truth telling in people who experienced no or only seldom punishment for deception in their learning history.

Registered Replication Report on Srull and Wyer (1979)

Srull and Wyer (1979) demonstrated that exposing participants to more hostility-related stimuli caused them subsequently to interpret ambiguous behaviors as more hostile. In their Experiment 1, participants descrambled sets of words to form sentences. In one condition, 80% of the descrambled sentences described hostile behaviors, and in another condition, 20% described hostile behaviors. Following the descrambling task, all participants read a vignette about a man named Donald who behaved in an ambiguously hostile manner and then rated him on a set of personality traits. Next, participants rated the hostility of various ambiguously hostile behaviors (all ratings on scales from 0 to 10). Participants who descrambled mostly hostile sentences rated Donald and the ambiguous behaviors as approximately 3 scale points more hostile than did those who descrambled mostly neutral sentences. This Registered Replication Report describes the results of 26 independent replications ( N = 7,373 in the total sample; k = 22 labs and N = 5,610 in the primary analyses) of Srull and Wyer’s Experiment 1, each of which followed a preregistered and vetted protocol. A random-effects meta-analysis showed that the protagonist was seen as 0.08 scale points more hostile when participants were primed with 80% hostile sentences than when they were primed with 20% hostile sentences (95% confidence interval, CI = [0.004, 0.16]). The ambiguously hostile behaviors were seen as 0.08 points less hostile when participants were primed with 80% hostile sentences than when they were primed with 20% hostile sentences (95% CI = [−0.18, 0.01]). Although the confidence interval for one outcome excluded zero and the observed effect was in the predicted direction, these results suggest that the currently used methods do not produce an assimilative priming effect that is practically and routinely detectable.

Registered Replication Report on Mazar, Amir, and Ariely (2008)

The self-concept maintenance theory holds that many people will cheat in order to maximize self-profit, but only to the extent that they can do so while maintaining a positive self-concept. Mazar, Amir, and Ariely (2008, Experiment 1) gave participants an opportunity and incentive to cheat on a problem-solving task. Prior to that task, participants either recalled the Ten Commandments (a moral reminder) or recalled 10 books they had read in high school (a neutral task). Results were consistent with the self-concept maintenance theory. When given the opportunity to cheat, participants given the moral-reminder priming task reported solving 1.45 fewer matrices than did those given a neutral prime (Cohen’s d = 0.48); moral reminders reduced cheating. Mazar et al.’s article is among the most cited in deception research, but their Experiment 1 has not been replicated directly. This Registered Replication Report describes the aggregated result of 25 direct replications (total N = 5,786), all of which followed the same preregistered protocol. In the primary meta-analysis (19 replications, total n = 4,674), participants who were given an opportunity to cheat reported solving 0.11 more matrices if they were given a moral reminder than if they were given a neutral reminder (95% confidence interval = [−0.09, 0.31]). This small effect was numerically in the opposite direction of the effect observed in the original study (Cohen’s d = −0.04).

Using more different and more familiar targets improves the detection of concealed information

When embedded among a number of plausible irrelevant options, the presentation of critical (e.g., crime-related or autobiographical) information is associated with a marked increase in response time (RT). This RT effect crucially depends on the inclusion of a target/non-target discrimination task with targets being a dedicated set of items that require a unique response (press YES; for all other items press NO). Targets may be essential because they share a feature - familiarity - with the critical items. Whereas irrelevant items have not been encountered before, critical items are known from the event or the facts of the investigation. Target items are usually learned before the test, and thereby made familiar to the participants. Hence, familiarity-based responding needs to be inhibited on the critical items and may therefore explain the RT increase on the critical items. This leads to the hypothesis that the more participants rely on familiarity, the more pronounced the RT increase on critical items may be. We explored two ways to increase familiarity-based responding: (1) Increasing the number of different target items, and (2) using familiar targets. In two web-based studies (n = 357 and n = 499), both the number of different targets and the use of familiar targets facilitated concealed information detection. The effect of the number of different targets was small yet consistent across both studies, the effect of target familiarity was large in both studies. Our results support the role of familiarity-based responding in the Concealed Information Test and point to ways on how to improve validity of the Concealed Information Test.

Validity of the Reaction Time Concealed Information Test in a Prison Sample

Detecting whether a suspect possesses incriminating (e.g., crime-related) information can provide valuable decision aids in court. To this means, the Concealed Information Test (CIT) has been developed and is currently applied on a regular basis in Japan. But whereas research has revealed a high validity of the CIT in student and normal populations, research investigating its validity in forensic samples in scarce. This applies even more to the reaction time-based CIT (RT-CIT), where no such research is available so far. The current study tested the application of the RT-CIT for an imaginary mock crime scenario both in a sample of prisoners (n = 27) and a matched control group (n = 25). Results revealed a high validity of the RT-CIT for discriminating between crime-related and crime-unrelated information, visible in medium to very high effect sizes for error rates and reaction times. Interestingly, in accordance with theories that criminal offenders may have worse response inhibition capacities and that response inhibition plays a crucial role in the RT-CIT, CIT-effects in the error rates were even elevated in the prisoners compared to the control group. No support for this hypothesis could, however, be found in reaction time CIT-effects. Also, performance in a standard Stroop task, that was conducted to measure executive functioning, did not differ between both groups and no correlation was found between Stroop task performance and performance in the RT-CIT. Despite frequently raised concerns that the RT-CIT may not be applicable in non-student and forensic populations, our results thereby do suggest that such a use may be possible and that effects seem to be quite large. Future research should build up on these findings by increasing the realism of the crime and interrogation situation and by further investigating the replicability and the theoretical substantiation of increased effects in non-student and forensic samples.

Manipulating item proportion and deception reveals crucial dissociation between behavioral, autonomic, and neural indices of concealed information

Developed as an alternative to traditional deception detection methods, the concealed information test (CIT) assesses recognition of critical (e.g., crime-relevant) "probes." Most often, recognition has been measured as enhanced skin conductance responses (SCRs) to probes compared to irrelevant foils (CIT effect). More recently, also differentially enlarged reaction times (RTs) and increased neural activity in the bilateral inferior frontal gyrus, the right middle frontal gyrus, and the right temporo-parietal junction have been observed. The aims of the current functional magnetic resonance imaging (fMRI) study were to (1) investigate the boundary conditions of the CIT effects in all three measures and thereby (2) gain more insight into the relative contribution of two mechanisms underlying enhanced responding to concealed information (i.e., orienting versus response inhibition). Therefore, we manipulated the proportion of probe versus irrelevant items, and whether suspects were instructed to actively deny recognition of probe knowledge (i.e., deceive) during the test. Results revealed that whereas overt deception was not necessary for the SCR CIT effect, it was crucial for the RT and the fMRI-based CIT effects. The proportion manipulation enhanced the CIT effect in all three measures. The results indicate that different mental processes might underlie the response pattern in the CIT. While skin conductance responding to concealed information may best be explained by orienting theory, it seems that response inhibition drives RT and blood oxygen level dependent responding to concealed information.

Reaction time measures in deception research: comparing the effects of irrelevant and relevant stimulus-response compatibility

Evidence regarding the validity of reaction time (RT) measures in deception research is mixed. One possible reason for this inconsistency is that structurally different RT paradigms have been used. The aim of this study was to experimentally investigate whether structural differences between RT tasks are related to how effective those tasks are for capturing deception. We achieved this aim by comparing the effectiveness of relevant and irrelevant stimulus-response compatibility (SRC) tasks. We also investigated whether an intended but not yet completed mock crime could be assessed with both tasks. Results showed (1) a larger compatibility effect in the relevant SRC task compared to the irrelevant SRC task, (2) for both the completed and the intended crime. These results were replicated in a second experiment in which a semantic feature (instead of color) was used as critical response feature in the irrelevant SRC task. The findings support the idea that a structural analysis of deception tasks helps to identify RT measures that produce robust group effects, and that strong compatibility effects for both enacted crimes as well as merely intended crimes can be found with RT measures that are based on the manipulation of relevant SRC.

Learning to lie: effects of practice on the cognitive cost of lying

Cognitive theories on deception posit that lying requires more cognitive resources than telling the truth. In line with this idea, it has been demonstrated that deceptive responses are typically associated with increased response times and higher error rates compared to truthful responses. Although the cognitive cost of lying has been assumed to be resistant to practice, it has recently been shown that people who are trained to lie can reduce this cost. In the present study (n = 42), we further explored the effects of practice on one’s ability to lie by manipulating the proportions of lie and truth-trials in a Sheffield lie test across three phases: Baseline (50% lie, 50% truth), Training (frequent-lie group: 75% lie, 25% truth; control group: 50% lie, 50% truth; and frequent-truth group: 25% lie, 75% truth), and Test (50% lie, 50% truth). The results showed that lying became easier while participants were trained to lie more often and that lying became more difficult while participants were trained to tell the truth more often. Furthermore, these effects did carry over to the test phase, but only for the specific items that were used for the training manipulation. Hence, our study confirms that relatively little practice is enough to alter the cognitive cost of lying, although this effect does not persist over time for non-practiced items.

Hyporeactivity of ventral striatum towards incentive stimuli in unmedicated depressed patients normalizes after treatment with escitalopram

Major Depressive Disorder (MDD) involves deficits in the reward system. While neuroimaging studies have focused on affective stimulus processing, few investigations have directly addressed deficits in the anticipation of incentives. We examined neural responses during gain and loss anticipation in patients with MDD before and after treatment with a selective serotonin reuptake inhibitor (SSRI). Fifteen adults with MDD and 15 healthy participants, matched for age, verbal IQ and smoking habits, were investigated in a functional magnetic resonance imaging (fMRI) study using a monetary incentive delay task. Patients were scanned drug-free and after 6 weeks of open-label treatment with escitalopram; controls were scanned twice at corresponding time points. We compared the blood oxygenation level dependent (BOLD) response during the anticipation of gain and loss with a neutral condition. A repeated measures ANOVA was calculated to identify effects of group (MDD vs. controls), time (first vs. second scan) and group-by-time interaction. Severity of depression was measured with the Hamilton Rating Scale of Depression and the Beck Depression Inventory. MDD patients showed significantly less ventral striatal activation during anticipation of gain and loss compared with controls before, but not after, treatment. There was a significant group-by-time interaction during anticipation of loss in the left ventral striatum due to a signal increase in patients after treatment. Ventral striatal hyporesponsiveness was associated with the severity of depression and in particular anhedonic symptoms. These findings suggest that MDD patients show ventral striatal hyporesponsiveness during incentive cue processing, which normalizes after successful treatment.