Abstract
Actively detecting deception requires (a) gathering information for fact-checking the communication content, (b) strategically prompting deception cues, and (c) encouraging honest admissions and discouraging continued deceit. Most deception-detection research, active or otherwise, finds that people are only slightly better than chance at correctly distinguishing truth from lies. Poor accuracy stems from a lack of reliable deception cues that hold across people and situations. Consequently, basing lie detection on deception cues is prone to error. However, some approaches to active deception detection yield higher accuracy than passive observation. Not all active approaches are advantageous. Mere interaction and mere question-asking produce outcomes similar to passive observation. Evidence-based and confession-solicitation approaches can be highly effective: for example, strategic use of evidence (SUE) and the content in context approach.
Tweet
Scientific evidence shows that lie detection improves with active approaches that use evidence strategically.
Key Points
Actively detecting deception is best achieved by gathering information for fact-checking the communication content and encouraging honest admissions.
Passive observation approaches that use cues or demeanor are little better than chance.
Trained experts can perform extremely well at lie detection.
Introduction
This article is about how to most accurately detect lies and deception. Deception exists, although it may be less prevalent than sometimes thought (Serota, Levine, & Boster, 2010). The presumption here is that it is important to correctly distinguish between honest and deception communication. The theoretical basis for this article is Truth-Default Theory (Levine, 2014), but the focus is more on the practical application of solid scientific evidence than on theory.
The title of this article implies a distinction between active and passive lie detection. This distinction can matter very much. But, at least with respect to accuracy in deception detection, passive and active approaches do not necessarily lead to different outcomes. Active methods can improve deception-detection accuracy considerably, but they can also lead to systematic errors and reduced accuracy. This article reviews the scientific evidence evaluating the efficacy of various approaches to deception detection and discusses the implications of those findings for policy and practice. The emphasis is on active deception detection that improves accuracy without technological equipment such as the polygraph or brain scans.
Active and Passive Approaches of Deception Detection
The distinction between active and passive deception detection, as used here, parallels earlier distinctions among passive, active, and interactive uncertainty-reduction strategies (Berger, 1979; Berger, Gardner, Parks, Schulman, & Miller, 1976). In this view, people typically try to reduce uncertainty about the people with whom they interact. People want to understand why people do what they do and anticipate what people will do in the future. That is, people want to predict and explain each other’s behaviors. Attempting to assess the honesty of another person’s communication can be a specific case of uncertainty reduction.
Reducing uncertainty requires gathering information. People use a variety of methods to reduce uncertainty, including three broad categories: passive strategies, active strategies, and interactive strategies. Passive strategies involve observing behavior unobtrusively. We can learn much from watching someone. Watching videotaped communication, reading an interview transcript, or watching an interview as an observer would be examples of passive observation. Active strategies go beyond mere observation but do not involve direct interaction. Asking other people (third parties) is a common active uncertainty-reduction strategy. Background checks and other types of evidence-gathering also fall under the category of active strategies. Interactive strategies involve direct interaction between the target person and the person trying to reduce uncertainty. Examples include directly asking questions or self-disclosing information about the self to trigger reciprocation. One effective way to detect deception through interaction: Ask a question where you already know the answer.
Distinctions among passive, active, and interactive strategies have useful application in deception detection. For the present purpose, however, active and interactive strategies can be lumped together under the label “active.” (In contrast, “passive” strategies for detecting deception mean observing—watching, listening, reading—some communication and assessing whether it is honest or deceptive.) Active strategies involve going beyond mere observation to gather evidence, to interact with a possible liar, or both. Evidence-based active strategies involve collecting other information to compare with what is said. Obviously, statements that contradict known or probable facts are suspect. Active strategies involving interaction typically involve either prompting behaviors useful in honesty assessment or in encouraging honest information including confessions. Evidence-based and interaction-based strategies can be used in combination, for example, by confronting a person with evidence (e.g., the strategic use of evidence [SUE] approach discussed later).
Beyond whether passively observed or actively prompted, the target’s behaviors can be usefully subdivided into communication content and cues/demeanor. Communication content refers to what is said, whereas demeanor refers more to how something is said. Demeanor is the overall impression a person gives off, whereas cues are specific behaviors (e.g., amount of eye contact, tone of voice) that, in combination, create a demeanor.
The Science of Deception Detection
Much of the research on deception detection has focused on two related issues. The first issue involves documenting observable behaviors that probabilistically signal deception. The second focus in on people’s ability to accurately distinguish truths for lies. Clearly, the ability to distinguish truths from lies depends on the existence of cues or information that differentiates honest and deceptive communication. However, the existence of useful cues or information does not mean that cues are effectively used to form a correct inference about honesty. Valid clues can be ignored or interpreted improperly. Both issues—cues and accuracy—appear in work on both passive and active approaches.
Passive Deception Detection
Most research on deception-detection accuracy involves tasks that fall squarely within the scope of passive deception detection. Most often would-be lie detectors experiments rate videotapes of truths and lies. But, just about every medium and communication mode imaginable has been studied (e.g., text on paper, text from online chat, social media pages, email, phone conversations, television, radio, live face-to-face interaction etc.). As it turns out, medium does not matter much. The subject population (students, working adults, police, etc.) does not make much difference either. Until recently, the findings were always the same. People were not much better or worse than “slightly above chance” at distinguishing truths from lie.
A meta-analysis of deception-detection research (Bond & DePaulo, 2006) reported that people were, on average, just below 54% correct when trying to distinguish between truths and lies. An accuracy of 50% could be obtained by chance, and 54% average was statistically above chance. Therefore, compared with chance, accuracy was “highly significant.” But, that said, accuracy is nevertheless much closer to chance than it is to perfection. Thus, although pure chance can be rejected with scientific confidence, from a practical application point of view, accuracy is dismal.
Their most informative but underappreciated finding, however, is the small variation around the 54% average. What this means is that the vast majority of studies provide remarkably consistent findings. Unlike most social scientific findings where the average is a statistical abstraction and most people aren’t “average,” in deception-detection research, most findings are not meaningfully different from the across-study average. Of the hundreds of accuracy findings in the literature, the vast majority fall within 44% and 65%.
Because most past deception-detection accuracy findings fall within this narrow range, these values provide reasonable benchmarks for assessing findings in reference to the larger literature as a whole. Findings within this range are nothing new and are not so different from the “slightly better than chance” across-study average to be atypical or attention grabbing. Thus, a finding of 61% is above the 54% average, but not outside the natural variation that exists around the mean. In contrast, findings in the mid-40% range and below suggest systematic errors and atypically poor performance, whereas findings 65% and above qualify as usually high. Findings outside the 44% and 65% range that replicate are not mere anomalies and reflect something newsworthy. Those are the findings that we will focus on.
Findings from the passive deception-detection accuracy studies summarized by Bond and DePaulo (2006) are listed in Table 1 along with findings from more recent passive and active approaches. As can be seen in the table, “slightly better than chance” accuracy holds across different communication media, regardless of sender motivation, regardless of whether the truths and lies are spontaneous or preplanned, regardless of whether or not judges have access to baseline honest answers for comparison, and regardless of whether or not the judges have professional experience in lie detection. Only two exceptions to poor accuracy with passive observations appear to replicate: “content in context” (and the similar idea of situational familiarity) and passive observation of effective expert questioning. Both of these exceptions are too new to have been included in the 2006 meta-analysis, and both are discussed later because they straddle the distinction between passive and active detection.
Passive and Active Deception Detection Accuracy.
Expert samples only.
The other major focus of deception research has been on behavioral cues that might signal lying, and the results of cue studies help explain the poor accuracy in passive detection studies. The cue findings are summarized nicely in one meta-analysis (DePaulo et al., 2003), and the link between cues and accuracy is summarized in another (Hartwig & Bond, 2011; see also Bond, Levine, & Hartwig, in press; Levine et al., 2011). Together, these lead to the following scientifically sound conclusions:
Most of the behavioral cues that have been investigated show little or no differences in occurrence between truths and lies (DePaulo et al., 2003). That is, no scientific evidence supports the utility of most cues as indicating honesty or deception. This is true for cues linked with emotions, arousal, cognitive effort, and strategic efforts to appear credible.
For the cues that have statistically significant differences, the differences tend to be too small to be of much practical value in actual lie detection (DePaulo et al., 2003). Liars differ from other liars, and honest communicators differ from other honest communicators, much more than liars differ from honest communicators.
Cue findings are inconsistent across studies. Further analysis typically fails to resolve the heterogeneity (DePaulo et al., 2003). Generalizing about the utility of cues is not scientifically defensible.
The few substantial differences diminish over time as research accumulates. That is, a literature-wide trend across various cues shows evidence get weaker as more is produced. Findings don’t replicate well and instead regress toward no difference (Bond et al., in press).
Poor accuracy in deception detection stems from a lack of useful cues to deceit, rather than judges’ failure to focus on useful cues. That is, poor passive lie detection is not a result of judges paying attention to the wrong cues. Instead, the problem is a lack of a valid signal (Hartwig & Bond, 2011).
To summarize, with the exception of a few recent and informative findings, passive lie detection is poor. Passive lie detection is better than chance, but not by much. Poor performance in passive deception tasks is exceptionally well documented. The reason for this poor performance is a lack of reliable behavioral differences between honest communicators and liars. Training judges in cues produces only small gains (Hauch, Sporer, Michael, & Meissner, 2014), does not reduce false positive rates (Hauch et al., 2014), and produces minimal improvements over placebo controls (Levine, Feeley, McCornack, Harms, & Hughes, 2005). Consequently, deception detection based on passive observation of cues and demeanor is inaccurate.
Content in Context and Situational Familiarity
There are exceptions to the “slightly better than chance” accuracy typical of passive deception detection. One exception is when both the communication content is informative, and the people judging the message have the contextual knowledge to understand the content. Content, as noted, means what is said rather than how it is said. Understanding what is said often requires knowing the context in which something is said. Familiarity with context and situation can make content informative, but content taken out of context or without context can be misleading. A second exception, observation of expert questioning, is discussed later.
As an example of the importance of context, imagine watching a videotaped interview between a police officer and a young man and trying to decide if the young man is honest or lying. Without background (context), what the young man says is difficult to interpret, and honesty assessments are of necessity based on what can be observed (the young man’s demeanor). Does the young man seem nervous? Is he being evasive, or does he seem cooperative? The research reviewed previously shows that deception judgments based on this sort of impression are not very accurate. Now suppose watching the interview with a good understanding of the context. You know whether the young man is a suspect, a witness, or a victim. You know the nature of the alleged crime, the basic crime facts, and you know what was reported by other witnesses. Knowing about the context, listening carefully to what is said becomes more useful in assessing honesty.
Providing people with a bit of useful context can improve accuracy (Blair, Levine, & Shaw, 2010). Participants (either college students or government agents) watched videotapes of truths and lies in three different situations. Some of the tapes were of students being questioned about suspected cheating in an experiment, some tapes were of students being interviewed about a mock crime, and some of the tapes were of guilty or innocent suspects in an actual criminal investigation. The participants making the truth–lie determinations were either provided with a little context information (e.g., a few crime facts known by the investigator at the time of the interview) or no context information. Across several experiments, in the lacks-context conditions, accuracy was, on average, only slightly better than chance. Average accuracy in the context-provide condition, however, was 75%.
Content in context is similar to the idea of “situational familiarity.” College student research participants viewed interviews of other students offering truthful or deceptive opinions regarding controversial issues at their own university or an out-of-state university (Levine & McCornack, 2001). When subjects judged the opinions about issues on their campus, accuracy improved from near chance to 69%. These findings replicated several times with accuracy ranging from 57% to 72% (Reinhard, Sporer, & Scharmach, 2013; Reinhard, Sporer, Scharmach, & Marksteiner, 2011). Thus, when communication content appears a familiar context, and when people can rely on content rather than just cues, passive deception-detection accuracy improves, sometimes substantially. Of course, content in context and situational familiarity can be actively sought by people seeking to detect lies, making these active methods of lie detection.
Finally, verbal content training improves passive lie detection substantially. Hauch et al. (2014) report that content trained improves accuracy relative to controls gu = .65. In comparison, the effects for providing feedback (gu = .19), cue training (gu = .28), and multiple approaches (gu = .34) were less effective.
Active Deception Detection
Although not fully apparent at the time, a critical turning in the understanding of accurate and inaccurate lie detection was the publication by Park, Levine, McCornack, Morrison, and Ferrerra (2002) of an article titled “How People Really Detect Lies.” Instead of conducting the typical passive deception-detection experiment, Park et al. simply asked their research participants to recall a successfully detected lie and describe how it was that the lie was uncovered. The vast majority of the accounts of detected lies involved either the use of evidence, the liar confessing the lie, or both. Evidence came from third parties, from physical evidence, or the lie detector’s prior knowledge. Confessions were sometimes solicited through questioning, but were also sometimes spontaneous, or the result of the liar forgetting that he or she had lied and coming clean by accident. Sometimes evidence was used to solicit a confession. Many lies were detected well after the fact. Only 2% of reported lies were detected in real time based on cues.
Poor accuracy in deception detect experiments may be because the lie detection task in the lab is not how people really detect lies (Park et al., 2002). In lie detection experiments, the motivation to lie is either irrelevant or constant across potential liars. No useful evidence is available. Liars never confess. Looking back at Park et al., the big successes in improved lie detection in recent research were foretold by their results. Motives, evidence, and confessions play critical roles. Active approaches that consider motives, seek to apply evidence, and prompt honest confessions produce much improved accuracy. Active approaches based only on deception cues produce accuracy levels similar to passive approaches or worse.
When Active Isn’t Better
Not all active approaches to deception detection have efficacy. For example, judgments made after direct interaction with a potential liar produce similar outcomes to passive observation absent any interaction (52.8% vs. 52.6%; Bond & DePaulo, 2006). Similarly, in a finding called the probing effect, the presence or absence of probing questions is unrelated to accuracy (Levine & McCornack, 2001). The probing effect includes both asking questions and merely hearing questions and answers; the effect is not moderated by whether or not the person making the judgment is the question asker. Instead, senders who are questioned are more likely to be believed regardless of their actual honesty (Levine & McCornack, 2001). Furthermore, the mere duration of the interview is unrelated to accuracy. Studies (Levine, Clare, et al., 2014; Levine, Shaw, & Shulman, 2010) found no correlation between interview duration and detection accuracy. Because cues have little utility and push accuracy down toward chance, more cues does not translate into higher accuracy. Thus, mere interaction and mere questioning–asking produce little if any improvement over passive observation.
Much recent research centers on behavioral cues associated with cognitive load (Vrij, Hope, & Fisher, 2014). The cognitive load approach presumes that it is more cognitively effortful to lie than to tell the truth. For example, Vrij and Granhag (2012) contend that “There is overwhelming evidence that lying is cognitively more difficult than telling the truth” (p. 112). The cognitive load approach involves adding additional load to a potential liar. Vrij and Granhag (2012) explain the rationale for imposing load:
Investigators can exploit the differences in cognitive load that liars and truth tellers experience. If lying requires more cognitive resources than truth telling, liars will have fewer cognitive resources left over. If cognitive demand is further raised, which could be achieved by making additional requests, liars may not be as good as truth tellers in coping with these additional requests. (p. 113)
Examples of this approach include requiring interviewees to provide a narrative in reverse chronological order, asking unanticipated questions, and instructing communicators to maintain eye contact. Recent results suggest that instilling load may have improved efficacy when used in conjunction deception cue training (Evans, Michael, Meissner, & Brandon, 2013).
Negative Utility
Some approaches to questioning can go terribly wrong and can actually produce accuracy that is worse than chance, especially when viewed by law enforcement professionals (Levine, Blair, & Clare, 2014; Levine, Kim, & Blair, 2010). The term “negative utility” describes cues or information that systematically lead to incorrect inferences. In these studies, interviewees were questioned with ineffective questions, and the interviewers were shown to law enforcement and security professionals. Accuracy levels well below change were observed (21.5% to 41.9%; Levine, Blair, & Clare, 2014; Levine, Kim, & Blair, 2010). Presumably, questions can be asked in such a way as to make honest individuals nervous and uncertain, and professionals are likely to incorrectly infer that the responses to the questions reflect dishonesty rather than a poorly designed question that is difficult for an honest person to answer convincingly.
Strategic and Expert Questioning Done Right
Although asking questions does not always improve accuracy, and although asking the wrong questions can actually make the situation worse, credible new scientific evidence suggests that active questioning can, when done the right way, yield much improved outcomes: for example, accuracy of 68% with short 20-s video clips of questions and answers (44% accuracy in controls; Levine, Shaw, & Shulman, 2010) and with an even better questioning approach, accuracies of 71.2% to 77.5% across 5 different samples of judges that included college students, local law enforcement officers, and federal agents (Levine, Blair, & Clare, 2014). Each interview lasted only about 2.5 min, yet accuracy was consistently above 70%.
A specific questioning strategy that can be highly effective is the SUE approach. The idea is that when questioning a potential liar, it is better to initially withhold known evidence to see whether statements contradict the evidence. Late disclosure of evidence, compared with early disclosure, improved accuracy from 43% to 62% (Hartwig, Granhag, Stromwall, & Vrij, 2005). Accuracy further improved to an impressive 85% when the questioner-judges were trained in the SUE. Strategically timing and framing evidence disclosure can maximize accuracy (e.g., see Granhag, Stromwall, Willen, & Hartwig, 2013).
Federal agents interviewed honest and cheating students to assess whether they were honest or not about whether nor not they or a partner had cheated on a task (Levine, Clare, et al., 2014). Across five agents who interviewed 89 potential cheaters, the agents correctly distinguished truthful and deceptive denials and admissions with 97.8% accuracy. Students later shown a sample of the expert interviews obtained 93.6% accuracy. Thus, expert interviewing can be highly effective.
Research, however, has not always found high accuracy from expert questioning. A similar sample of federal agents who questioned potential cheaters in a similar situation obtained accuracy of 59% (Dunbar et al., 2013), which was not significantly different from the meta-analysis average of 54%. Two key elements of the research designs may explain this discrepancy in findings. First, Dunbar et al.’s experts followed a script, whereas Levine’s experts were not bound to specific questions. Second, although Dunbar et al. tallied confessions, confessors were excluded from accuracy calculations. Levine et al. counted honest confessions that were believed as correct. Scripting alone, however, is not sufficient to account for the differences. An experienced police officer interviewed student mock crime suspects unscripted, and accuracy was just 56.7% (Hartwig, Granhag, Stromwall, & Vrij, 2004). Thus, experts can be, but are not necessary, accurate at lie detection.
In summary, recent evidence finds that deception detection can be improved through some active deception methods. Improved accuracy is obtained though communication content that can be fact checked or at least assessed for plausibility, and by persuading liars to confess their dishonesty and tell the truth. When expert interviewers are familiar with the context, and when they are allowed to solicit confessions as a means of lie detection, expert questions can be highly effective. This evidence has practical and policy implications.
Implications and Conclusion
Ten years ago, the best available scientific evidence overwhelmingly supported the claim that people were invariably poor lie detectors. Given that understanding, the best pragmatic advice was twofold. People needed to understand the fallibility of human lie detection, so they might avoid faulty inferences and false confidence. Without the aid of technology, human lie detection was little better than a coin-flip. People needed to know this because research showed again and again that people thought they were better lie detectors then they actually were. At the time, the best hope for scientifically valid lie detection seemed to be some new technology. Brain scans seemed most promising, but several new technologies were being tested, and the use of the polygraph was still being refined and improved.
But, that was then. Today, the “slightly better than chance” accuracy that was once an unwelcome but undeniable fact appears to hold only under some conditions, and scientifically sound evidence supports approaches that yield much more accurate lie detection. The research is still in its infancy, but enough findings have now replicated enough times to know that the old understanding is giving way to a new view. The most promising approach these days uses improved interviewing and interrogation techniques focusing on communication content and persuasion.
The best available scientific evidence now suggests that accurate human lie detection involves communication content (what is said), rather than demeanor (how it is said) or cues. Evidence-based approaches and approaches that solicit honest confessions consistently produce the highest levels of accuracy. Approaches that allow content to be assessed for plausibility such as content in context and situational familiarity also lead to improved outcomes. As an example, doing background checks when hiring new employees, security screenings, and the like, and providing that information to the interviewers/screeners is likely to yield substantial improvement over interviews or behavioral observation alone. In the law enforcement context, good detective work is needed to understand crime facts; interviews and interrogations yield better outcomes when answers are understood in light of facts. But, this advice isn’t just for law enforcement and the intelligence community. Anyone can do a little “detective work” especially in the age of Google, social networks, and the Internet. If assessing the honesty of some communication is important, doing a little homework can go a long way in deception prevention and detection. Paying attentions to what is said in light of what can be realistically known is the key.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.