RESEARCH DOSSIER
Deception Detection and Credibility Judgment
Abstract. Deception detection examines how observers, interviewers, institutions, and computational systems distinguish intentional misrepresentation from truthful communication. Credibility judgment evaluates the epistemic warrant of a person, account, claim, or source, drawing on both communication and independent evidence. This monograph develops an interdisciplinary critical narrative synthesis of experimental deception research, social cognition, psycholinguistics, forensic interviewing, signal detection, psychophysiology, and emerging automated assessments. It reconstructs the transition from presumed behavioral signs of lying to probabilistic judgments about information, verifiability, and interview-generated evidence. Meta-analytic findings show that unaided human discrimination is modest and that behavioral cues are often weak; studies of strategically elicited evidence offer promising but context-dependent improvements. The analysis separates sincerity, accuracy, memory, and corroboration; treats error costs, ground truth, base rates, cultural variation, and procedural fairness as constitutive of credibility assessment; and examines the limits of physiological and artificial-intelligence-based approaches. The synthesis identifies unresolved problems in ecological validity, selective ground-truth verification, dependent datasets, independent field replication, and the consequences of false accusations. It reports neither new experimental data nor a systematic-review protocol.
Keywords: ; forensic psychology
Introduction and Scientific Object
The central scientific problem of deception detection is inferential: an evaluator observes words, conduct, or records but seeks to determine whether a communicator intentionally created or sustained a misleading understanding. Intent is not directly visible in a facial expression or speech hesitation. A truthful statement can be mistaken; a fabricated statement can contain many true details; and a misleading account can consist entirely of literally accurate sentences. Scientific assessment consequently depends on specifying what is to be inferred, the information available to the evaluator, and the independent means by which an inference may be evaluated.
Credibility judgment concerns a wider family of questions. An investigator might ask whether a witness reports an event accurately, whether a particular proposition is corroborated, whether an interviewee is withholding relevant information, or whether a source has demonstrated reliability across occasions. These judgments have different targets and ground-truth criteria. A witness can speak sincerely and misremember. Someone can accurately describe a real event while deceptively concealing another. A source can generally be dependable while making one consequential error. Collapsing these distinctions into a single impression of whether someone looks credible creates avoidable conceptual and methodological confusion.
This research dossier takes the observer's decision as its organizing unit. It examines the production of deceptive and truthful statements, the cues potentially available to a recipient, the conditions under which those cues become diagnostic, the recipient's decision policy, and the consequences of judgment. Its primary domain is forensic and legal psychology, with necessary connections to social cognition, memory, decision theory, communication science, psychometrics, and technological assessment. The scientific question is how justified differentiation can be achieved under uncertainty, with attention to both errors in calling a truth a lie and errors in accepting a lie as truthful.
The method is a structured critical narrative synthesis. Original experiments are interpreted alongside major meta-analyses, methodological papers, and institutional evidence assessments. This permits a detailed comparison of theories and research programs while preserving the distinction between data that have been independently evaluated and propositions that remain hypotheses. The synthesis does not claim systematic database coverage or pooled reanalysis. Its principal evidentiary anchors include DePaulo and colleagues' cue synthesis (2003) DePaulo (2003), Bond and DePaulo's accuracy meta-analysis (2006) Bond (2006), Hartwig and Bond's lens-model analysis (2011) Hartwig (2011), and Vrij, Hartwig, and Granhag's integrative review (2019) Vrij (2019).
Definitions: Deception, Truthfulness, Accuracy, and Credibility
A workable psychological definition of deception centers on a communicator's intentional management of another person's beliefs or understanding through misleading representation, omission, or contextual manipulation. A lie is a narrower form, frequently operationalized in experiments as a statement the communicator believes false and presents as true. Researchers must distinguish the communicator's belief from the objective truth value of the statement. When ground truth is known, the study can classify an account as factually correct or incorrect; when intent is known through experimental assignment, it can also examine deliberate misrepresentation. In ordinary investigations, both may initially be uncertain.
Sincerity designates the relationship between a statement and a speaker's belief or endorsed account. Accuracy designates correspondence with the relevant state of affairs. Completeness concerns whether pertinent information was omitted. Consistency concerns relationships among statements, records, and other evidence; it may be internal, temporal, interpersonal, or external. Credibility is the warranted confidence that a particular communicative claim or source merits for a particular purpose. These dimensions overlap without being interchangeable. An evaluator may rationally trust one part of testimony and challenge another because the corroboration differs across propositions.
An operational definition must also specify the unit of analysis. Some laboratory studies classify one brief statement; others compare full narratives, interview sessions, or speaker-level patterns. An evaluator who classifies ten statements from one person does not automatically have ten independent observations. Studies that infer whole-person dishonesty from a small set of recorded utterances are especially vulnerable to overgeneralization. Properly designed research distinguishes between statement-level, interaction-level, person-level, and institutional-level prediction.
The empirical literature also differentiates deception detection from misinformation detection. A false social-media claim may have been shared innocently, produced as satire, strategically fabricated, or generated by an automated system. Factuality classification can proceed by document and source verification even when deceptive intent remains unknown. The broader theoretical importance of this distinction increases in environments in which the producer of a message, the publisher, and the person who distributes it are separate actors. Levine's truth-default account (2022) Levine (2022) illustrates why the receiver's tendency to accept communication and the communicator's actual intention are analytically different variables.
Historical Development of Deception Research
The wish to identify falsehood through bodily behavior predates experimental psychology, but contemporary evidence must be evaluated through documented measures rather than presumed traditions. In the twentieth century, psychophysiological examinations connected deception claims to arousal, autonomic activity, and examination protocols. Subsequent social-psychological research focused on visible and audible behavior, asking whether the act of lying systematically changes eye movements, speech characteristics, expressions, posture, or spontaneous actions. The attraction of this program was methodological simplicity: if deceptive speakers emitted stable observable signals, a relatively passive observer could in principle detect them.
The late twentieth and early twenty-first centuries substantially changed that research picture. Instead of relying on anecdotes about suspicious appearance, investigators compared truthful and deceptive behavior under manipulated conditions, coded numerous cues, and aggregated effect sizes across experiments. DePaulo et al. (2003) DePaulo (2003) synthesized 1,338 estimates concerning 158 purported cues, reporting that many showed little or no dependable relationship to deception. This result shifted the explanatory burden toward whether any behavior could furnish sufficiently strong diagnostic evidence for decisions about individual people.
Research on observers evolved in parallel. Bond and DePaulo (2006) Bond (2006) aggregated deception judgments across hundreds of reports and nearly 25,000 judges. Later lens-model analyses distinguished available information from the information that judges use, testing whether incorrect cue beliefs or the inherent weakness of cues better explain limited discrimination. Hartwig and Bond (2011) Hartwig (2011) found that weak diagnostic signals were a major constraint, complicating the popular explanation that poor judgments arise entirely from a mistaken list of stereotyped behaviors.
A further development emphasized what an interviewer can cause a communicator to reveal. Rather than expecting concealed intent to leak automatically through the face, researchers developed methods for obtaining detailed accounts, introducing corroborating evidence at strategic points, and examining how different hypotheses predict the answer to a question. Hartwig et al. (2005) Hartwig (2005) and Hartwig et al. (2006) Hartwig (2006) are influential experimental examples. Vrij (2019) Vrij (2019) describes the historical movement from nonverbal observation to verbal content, from generalized arousal to cognitive strategies, and from passive observation to active information elicitation.
The history therefore describes several partially overlapping programs, each with different observable measures and validity requirements. Modern research integrates communication theory, cognitive psychology, memory science, and forensic practice while retaining unresolved disputes about the diagnostic value of particular procedures. The progression is methodological: increasingly explicit hypotheses, better control of ground truth, and greater concern with what a decision can legitimately infer from the obtained information.
Methods, Ground Truth, and Evidence Quality
In a typical experimental deception design, participants are assigned to produce honest or deceptive accounts about a controlled event, opinion, or activity. The design can establish the instructed status of the statement and manipulate incentives or interview conditions. Random assignment helps isolate causal effects of an intervention, but an instructed low-stakes lie differs in incentives, emotional consequences, rehearsal, and opportunity for strategic preparation from an account given during an actual investigation. The distinction between a controlled manipulation and its intended real-world construct must be explicit in any interpretation.
Field studies offer greater contextual realism but often encounter the harder problem of independently establishing whether the speaker was deceptive. A confession, accusation, conviction, or investigator's impression is an imperfect ground-truth label. If the label is partly generated by the process that a study seeks to validate, circularity arises. A purportedly accurate detector may merely learn a record of institutional decisions. Evaluations should disclose how each case was verified, what portion of the original population lacked verification, and whether false positives or false negatives were disproportionately missing.
Measurement proceeds along several levels. Behavioral coding specifies a cue and reliability among coders; linguistic analysis requires transparent segmentation and defined features; audience judgments require an identifiable response scale; and interventions require a stable comparison condition. Interrater agreement, coder blinding, stimulus assignment, and pre-registered analytic decisions help reduce avoidable bias. A detector trained and tested on data from the same speaker, event, or video source may benefit from hidden shared information even when the individual clips are separated into different training folds.
The strongest comparative studies distinguish development, internal validation, and external validation. Internal cross-validation can reveal overfitting within the sampled source, whereas genuinely new speakers, interviewers, institutions, languages, topics, and recording equipment test transportability. The same principle applies to interventions: a procedure shown effective among trained student interviewers and mock suspects must be evaluated independently before it is assumed effective in criminal inquiries. Camara et al. (2024) Camara (2024) report consequential differences among low- and high-stakes video datasets and warn against treating available datasets as sufficient for deployment.
A disciplined evidence assessment therefore records the research question, population, ground-truth criterion, sampling frame, stimulus characteristics, unit of independence, interview intervention, outcome metric, comparative baseline, and error costs. Meta-analysis provides valuable aggregation only when underlying studies have sufficiently comparable outcomes and limitations are carried through the synthesis. Expert assessment of a witness or automated classifier needs a separate validation record; the publication of an impressive accuracy number alone does not establish reliability for a consequential forensic decision.
Statistical Discrimination and Decision Thresholds
Deception judgments are classification decisions under uncertainty. Their performance depends on two partly separable properties: discrimination and response criterion. Discrimination expresses how effectively evidence distinguishes the conditions of interest; a response criterion expresses how much evidence an observer requires before assigning a label such as deceptive. An observer may be strongly biased toward believing people while still discriminating among statements better than chance, or may accuse many people without achieving meaningful discrimination.
Signal detection theory clarifies these relationships. In a binary design, sensitivity to lying may be represented by d-prime, which compares the standardized evidence distributions for deceptive and truthful conditions. Criterion measures a decision tendency; its movement can change the proportions of accepted truths, missed lies, false accusations, and correctly detected lies without improving sensitivity. Receiver-operating-characteristic curves examine performance across possible thresholds. Area under the curve summarizes ranking performance, but does not specify the error rate at the decision point an institution actually intends to use.
Prevalence changes the practical meaning of a positive result. For example, in an explicitly illustrative population of 1,000 communications with a 10% prevalence of deception, a hypothetical test with 80% sensitivity and 80% specificity would flag 80 of 100 deceptive cases and 180 of 900 truthful cases. Of the 260 positive classifications, approximately 31% would identify deception correctly. These figures are a mathematical illustration, not an estimate of performance by any published detector. They show why even a seemingly capable system can generate many false accusations in a low-prevalence setting.
The false-positive cost may be unusually high when credibility judgments shape criminal allegations, asylum determinations, employment decisions, or child-protection inquiries. False negatives can also carry serious costs. Institutional use therefore requires explicit specification of which errors matter, who bears them, and whether the purpose is triage, investigative questioning, corroboration, or final adjudication. A system designed to produce investigative leads is not validated as a basis for a finding of culpability.
Calibration is another distinct requirement. When a model or human assessor reports 80% confidence in an accusation, similarly rated cases should correspond to that outcome at approximately the stated rate in the target environment. Strong discrimination can coexist with poor calibration, and calibration may deteriorate when prevalence or populations change. Responsible credibility research reports sensitivity, specificity, predictive values at relevant base rates, threshold behavior, calibration, and uncertainty rather than presenting overall accuracy as a universal property of the evaluator.
The Human Accuracy Literature
The classic pooled estimate derives from Bond and DePaulo (2006) Bond (2006), who synthesized 206 documents involving 24,483 judges. Across their sampled conditions, people achieved approximately 54% correct truth-lie judgments. They correctly classified about 47% of lies and 61% of truths. The design principally involved real-time, unaided judgments rather than independent forensic corroboration, and the pooled value is a description of those researched tasks rather than a law governing every conversation.
The difference between correct classification of truths and lies is theoretically important. A truth-favoring response criterion may be socially adaptive in ordinary cooperation, while a lie-favoring criterion may emerge when suspicion is elevated by an experimental instruction or institutional setting. Accuracy therefore needs to be reported separately for both message types. A detector that calls everyone a liar could correctly identify every deceptive statement in a balanced sample while falsely accusing all honest speakers; its apparent success on one subscore would conceal total failure of discriminative usefulness.
The meta-analysis also found variation by presentation modality, including higher average accuracy when observers could hear rather than merely see speakers. This pattern supports attending to what is said and to the opportunity to evaluate content rather than assuming that visual access to demeanor supplies the decisive evidence. Yet hearing a statement is not equivalent to confirming its factual content. An eloquent account can be inaccurate, and a confused presentation can be sincere. Apparent credibility must be evaluated against task-specific, independently grounded information.
Subsequent investigations of individual differences have examined expertise, acquaintance, experience, and response strategies. The broad literature does not establish that routine professional exposure to suspects or witnesses creates a general-purpose ability to recognize deception across unfamiliar contexts. The more productive question is which specific professional practices obtain additional valid information and how those practices can be assessed in controlled, independently replicated studies.
The 54% statistic is often misused as a forecast for any individual interrogator, every cultural setting, or every kind of content. Its force lies in demonstrating the limited discriminative information available in the studied conventional judgment tasks. Hartwig and Bond's later lens-model synthesis (2011) Hartwig (2011) reinforces this interpretation: the difficulty of lie judgment is substantially connected to the weakness of the behavioral evidence being judged. Evaluation quality improves when a study asks what new evidence would separate a truthful account from the particular alternative hypothesis at issue.
Truth-Default, Suspicion, and Social Cognition
Levine's truth-default theory (2014) Levine (2014) argues that people ordinarily process communication within an expectation of sincerity, and that active suspicion arises when particular contextual triggers become salient. The framework explains why a person may engage in an extended conversation without consciously checking each sentence for possible falsehood. The default is a feature of human communicative organization, in which continuous explicit verification of every statement would be expensive and socially disruptive.
Experimental evidence directly addresses the distinction between unprompted social processing and prompted lie detection. Clare and Levine (2019) Clare (2019) examined verbalized thoughts in videotaped and face-to-face interactions. In their conditions, spontaneous, unprompted consideration of veracity appeared in fewer than one in ten coded thought listings. The experiment supports the existence of a truth-default in those settings, while the exact behavioral implications depend on the interaction, the plausibility of content, and the measurement procedure.
A prompted laboratory test makes veracity salient before the observer has any reason to question a speaker. That instruction changes the task relative to ordinary conversation. A forensic interviewer may begin from suspicion, an investigator may have external intelligence, and a friend may rely on shared history. These are differences in prior information and decision context rather than simple variations in individual skill. The same observed sentence may be assessed differently when an independently documented contradiction is available.
Bias and discrimination also require separation. Increased skepticism can produce more lie classifications while leaving actual sensitivity unchanged. Conversely, calibrated trust toward a consistently reliable source may be warranted, although no history removes the need to evaluate contradictory evidence. Levine (2022) Levine (2022) extends the theory to information environments in which belief congruence, social repetition, and source agreement influence acceptance of messages. These applications concern message acceptance as well as deliberate deception; they cannot be reduced to visual indicators of whether one speaker is lying.
Forensic credibility evaluation should therefore ask what triggered suspicion, which evidence was available before the interview, whether the evaluator was exposed to irrelevant allegations, and how a contrary hypothesis would be tested. This preserves the possibility of both believing warranted testimony and correcting an initially attractive but unsupported accusation.
Why Behavioral Cues Are Weak
A behavior can be associated with deception in a group comparison without being a reliable diagnostic marker for a particular person. DePaulo et al. (2003) DePaulo (2003) found that liars in some research conditions tended to be less forthcoming, less compelling, and more tense, but many investigated cues were weak or unrelated to deception. Their synthesis also identified moderators connected to motivation and the nature of the lie. A group-level average is consequently inadequate as an individual inference rule.
There are several reasons for overlap. Anxiety can arise from being falsely accused; rehearsed truth can sound polished; an unfamiliar question can disrupt a sincere person's speech; and a practiced deceiver can provide fluent detail. People also differ in expressive habits, speech conditions, fatigue, language proficiency, neurodevelopment, and social expectations. Many of these factors change the same behaviors investigators might label suspicious. The resulting distributions of visible conduct among truthful and deceptive speakers can overlap extensively.
Hartwig and Bond (2011) Hartwig (2011) used a lens-model approach to ask whether observers primarily select the wrong indicators. Their analyses showed substantial correspondence between cues associated with perceived deception and cues associated with actual deception, while indicating that actual cues often provide insufficient discriminative strength. The result challenges interventions that assume a short course correcting eye-contact myths will create strong detection performance.
The causal problem is equally consequential. A cue such as hesitation might be influenced by memory retrieval demands, conversational politeness, translation, or a speaker's attempt to formulate an accurate answer. Even when a difference is reproducible under one protocol, the mechanism must be tested across alternative explanations. If a behavior is caused by interview pressure rather than intentional misrepresentation, treating it as evidence of lying may convert a procedure-induced effect into an apparently independent accusation.
The appropriate evidentiary question is therefore whether a proposed cue adds incremental predictive validity beyond what is already known from the claim, context, and corroborating records, whether the cue retains validity in independent samples, and whether it yields acceptable false-accusation rates. This requirement follows from the scientific interpretation offered by Vrij, Hartwig, and Granhag (2019) Vrij (2019), whose review describes many apparent nonverbal signs as faint, unstable, and easily misconstrued.
Nonverbal Behavior, Emotion, and Microexpressions
Research on nonverbal behavior examines facial movement, gaze, blink patterns, gesture, posture, vocal pitch, timing, and interpersonal coordination. These measures concern observable behavior; their relation to deliberate misrepresentation must be established in each experimental context. The widespread claim that averting one's gaze reveals lying conflates cultural conventions, attention, anxiety, interactional demand, and deception. Some individuals maintain eye contact while deceiving, and truthful people often avert gaze while concentrating on memory or searching for words. A feature can support an impression of discomfort without providing sound evidence of a false statement.
Facial-expression research has contributed carefully measured findings while revealing important limits of generalization. In an experiment involving genuine and deliberately altered emotional displays, Porter and ten Brinke (2008) Porter (2008) analyzed frame-by-frame expressions and observed instances of inconsistent emotional signals. Their microexpression measure occurred in a minority of participants and a small proportion of displays. The experiment concerns the concealment and simulation of emotional reactions under controlled stimuli. Evidence that emotion can be incompletely concealed does not establish that a fleeting facial movement detects factual lying in forensic interviews.
Emotion, arousal, and deception also have different causal locations. An interviewee may feel fear because a truthful account is being doubted, embarrassment over a separate private matter, anger at institutional authority, or concern about consequences unrelated to the question. A deceiver may feel calm, or may experience pride and control. The relationship between emotional display and intent is mediated by situation and appraisal. Extracting an emotion label from the face therefore leaves a substantial inferential gap before the evaluator can conclude that deception occurred.
Vrij, Hartwig, and Granhag (2019) Vrij (2019) reviewed why apparently intuitive nonverbal signs perform poorly as lie criteria. Methodological progress requires longitudinal observation, multimodal coding, independent truth verification, and investigation of interactions among cues, while still checking whether the same cues appear in truthful but stressed speakers. Neither a facial-emotion recognition score nor a high interrater agreement in cue coding substitutes for veracity validation. Their evidentiary function remains a hypothesis to be tested rather than a general warrant for accusing a particular person.
Verbal Content and Narrative Structure
The content of a statement can contain information about the reported event, the person's knowledge, and the opportunities for independent testing. Verbal analysis examines specificity, temporal sequence, sensory description, contextual embedding, spontaneous corrections, and relationships between claims and available records. Compared with stereotyped interpretations of gestures, such features can be connected more directly to competing accounts of what happened. A statement that places an action at a specific location and time can be checked against independent records; a broad emotional impression generally cannot.
This advantage must be kept within its proper scope. Truth tellers differ in memory, fluency, and motivation to describe an event; liars may know genuine details and incorporate them into a misleading narrative. Detailed accounts may be fabricated or rehearsed. Sparse accounts may reflect stress, a long retention interval, limited language proficiency, or reluctance to discuss a traumatic experience. The existence of a linguistic difference in a group study does not demonstrate that a single detail count reliably diagnoses lying in an individual case.
A content-based approach gains scientific value when it describes an explicit mechanism. If a truth teller has directly experienced an event and a liar must invent or strategically select details, some differences in reported information may follow. Yet this prediction depends on whether the deceiver has access to real event details, whether the truth teller remembers them, whether the interview invites elaboration, and whether both parties can prepare. It also depends on how the researcher codes detail. Independent coding and preregistered analytic rules become important whenever an apparently small coding choice changes the classification outcome.
Vrij (2019) Vrij (2019) characterizes verbal and strategically elicited information as an important direction of contemporary lie-detection research. An evidence-centered application should compare verbal output with separately established knowledge and record how the information was obtained. Linguistic richness can guide further questions and corroboration, but credibility ultimately depends on the warrant supplied by the whole evidentiary record rather than on stylistic conformity to a presumed truthful narrative.
Criteria-Based Content Analysis and Reality Monitoring
Criteria-based content analysis, or CBCA, investigates qualitative attributes of statements within traditions of testimonial credibility assessment. Reality monitoring, or RM, draws from theories about differences between information derived from external experience and internally generated events. Both have been studied as methods for identifying content associated with genuine experiences rather than fabrication. Their core hypotheses concern processes of event encoding, memory retrieval, and narrative construction, which can be evaluated through comparative research.
The relevant systematic quantitative synthesis is Oberlader et al. (2016) Oberlader (2016), a meta-analysis of content-based approaches to distinguish true and fabricated statements. It establishes that these methods have an empirical literature and can produce group-level differences under studied conditions. Interpretation still depends on the mix of laboratory and field designs, the standards used to establish truth, the coding procedure, and whether the method is applied as a checklist, a probabilistic model, or part of a wider professional assessment.
An account of actual experience can be missing detail because of delay, age, fatigue, limited opportunity to perceive, or the demands of the interview. A fabricated account may borrow real autobiographical material. Interviewers can alter the available content through leading prompts or by supplying information to a witness. The validity of a criterion in one domain does not make the criterion equally diagnostic in every population. Where an assessment concerns a child, language learner, or vulnerable witness, the interviewer must also consider developmental language, retrieval conditions, and suggestibility.
The proper inference is probabilistic and conditional. A content-analysis score can contribute evidence about a specific type of statement under an adequately validated protocol; it cannot establish certainty about another person's experience or intention. Independent corroboration and examination of alternative explanations remain necessary. The same principle governs scientific interpretation of reality monitoring: the separation of remembered and imagined information is a research question whose answer varies with the event and retrieval conditions rather than a universal binary signature.
Cognitive Accounts and Interview-Generated Cues
Cognitive approaches propose that fabricating, maintaining, or adapting a deceptive account can place different demands on memory, attention, inhibition, and strategic monitoring from those associated with truthful retrieval. A deceiver may have to represent the true event, construct an alternative account, remember previous claims, anticipate the questioner's knowledge, and prevent contradictions. The mechanisms are plausible, but they are not uniformly more demanding than truthful reporting: difficult genuine memories can be cognitively taxing, while well-rehearsed lies can become fluent.
Blandón-Gitlin et al. (2014) Blandón-Gitlin (2014) examined the cognitive-load approach and emphasized the need to identify the underlying mechanisms rather than simply raising demands on an interviewee. Imposing a secondary task, reversing the order of events, or introducing an unexpected request may change both honest and deceptive performance. The scientific issue is whether a particular manipulation magnifies their difference and whether this benefit persists under ecologically credible conditions.
An important methodological distinction exists between provoking an informative answer and provoking general difficulty. A procedure that makes all respondents confused can produce more apparent inconsistencies without improving discrimination. In high-stakes situations, undue cognitive pressure may introduce fairness concerns and contaminate accurate memories. Investigative interview design should therefore seek questions that reveal knowledge and can generate independently testable claims, while keeping the process understandable and appropriately adapted to the person being interviewed.
Vrij, Fisher, and Blank (2017) Vrij (2017) meta-analyzed cognitive approaches to lie detection, providing a structured assessment of an intervention literature rather than one isolated demonstration. An effective methodological comparison needs an explicit control interview, independently defined deceptive and truthful conditions, observer blinding where possible, and reporting of both types of misclassification. Benefits measured in controlled conditions warrant further field validation before adoption as decisive forensic proof.
Strategic Use of Evidence
Strategic Use of Evidence, commonly abbreviated SUE, changes the relationship between the interviewer's prior knowledge and the interviewee's account. The interviewer identifies independently established facts and plans when and how to disclose them, allowing an interviewee's account to be compared with information the interviewee may not initially know is available. The technique uses statement–evidence consistency as a hypothesis-relevant outcome. Its value depends on genuine prior evidence, an appropriate interview procedure, and the opportunity for an innocent person to explain apparent discrepancies.
In a foundational experiment, Hartwig, Granhag, Strömwall, and Vrij (2005) Hartwig (2005) investigated detection through the strategic disclosure of evidence. The design tested how disclosure order influences the information available to the evaluator. Its significance is conceptual as much as empirical: the diagnostic signal is partly produced by the interview procedure rather than expected to appear spontaneously in a speaker's face.
A subsequent controlled study by Hartwig, Granhag, Strömwall, and Kronkvist (2006) Hartwig (2006) involved 82 police trainees and 82 mock suspects. The study reported 85.4% deception-classification accuracy among trained interviewers compared with 56.1% among untrained interviewers under its experimental conditions. These numbers identify the results of that particular training and simulation. They are not a general prediction of accuracy in real criminal cases, and the mechanism may depend on the particular evidence held by the interviewer and the suspect's strategy.
The critical insight is that contradictions acquire diagnostic value through their evidentiary meaning. An alleged inconsistency might arise from an inaccurate record, poor recall, misunderstanding of a question, an ambiguous time reference, or deliberately concealed information. Interviewers must distinguish contradiction from deception and document how both evidence and explanation were obtained. Independent case verification, cautious disclosure protocols, and preservation of original accounts help prevent post hoc selection of only those discrepancies that support an initial suspicion.
SUE is best understood as a structured information-gathering strategy whose promise requires independent replication, training fidelity, and prospective evaluation of both truth and lie outcomes. Its scientific contribution lies in making the interview an experimental opportunity to compare predictions derived from competing accounts rather than a performance in which a supposedly skilled interrogator reads guilt from demeanor.
Verifiable Details and External Corroboration
The verifiability approach asks which parts of an account can be checked through independent records, people, or events. Someone reporting a genuine activity may provide testable details about locations, transactions, witnesses, messages, or temporal sequences; a deceiver may prefer vague statements or details that are difficult to verify. This prediction becomes empirically meaningful only when the interview affords both groups a fair opportunity to supply checkable information and the investigators actually attempt verification.
Nahari, Vrij, and Fisher (2014) Nahari (2014) tested a verifiability approach in which countermeasures and strategic content influence differences between true and deceptive accounts. Their study illustrates that the diagnostic value of detail depends on how the speaker anticipates scrutiny. The method therefore investigates communicative strategy and available corroboration rather than merely counting adjectives or sensory words.
A verifiable detail is a claim with a potential checking procedure; it is not automatically a verified fact. Independently confirming that a person visited a location may strengthen that part of an account without demonstrating the truth of all stated intentions, actions, or accompanying claims. Conversely, the inability to check a detail may reflect missing records or privacy constraints rather than fabrication. Verification quality depends on source independence, temporal relevance, reliability of the record, and whether the source itself may have been contaminated by earlier statements.
Careful credibility judgment should preserve a map of claims, evidence, alternative explanations, and remaining gaps. It should distinguish confirmed, contradicted, untested, and presently untestable assertions. This approach permits partial conclusions and updates when further evidence arrives. Such proportionality matters especially where an interviewee makes a mixture of accurate and inaccurate claims, because global labels of honest or dishonest can suppress the finer structure of evidentiary support.
Training, Expertise, and Feedback
Training may improve deception classification by changing the information an evaluator obtains, the cues considered, or the decision criterion. These mechanisms have different consequences and require different outcome measures. Telling trainees that gaze aversion is a poor marker may reduce one stereotype but leave little additional valid information available. Training them to elicit detailed accounts and evaluate corroboration may alter the evidentiary task itself. Training that simply encourages more suspicion may increase true lie detections together with false accusations.
The meta-analysis by Hauch, Sporer, Michael, and Meissner (2016) Hauch (2016) included 30 control-group studies and reported a small-to-moderate overall improvement in accuracy, with more favorable effects for approaches emphasizing verbal content than those emphasizing nonverbal or paraverbal cues. The authors' estimates addressed the participating studies and their particular interventions. They do not justify advertising an observer as reliably able to identify deception after any generic training package.
Expertise should be evaluated as performance on well-defined tasks rather than occupational title. Investigators, judges, clinicians, border officers, and other professionals encounter different kinds of statements and may possess useful domain knowledge. That knowledge may permit an informed question or recognition of an independently established inconsistency, while leaving unaided reading of unfamiliar people highly uncertain. Ecologically realistic proficiency tests need representative cases, independent labels, controls for case difficulty, and calibration of expressed confidence.
Feedback poses a further problem. An evaluator who receives confirmation only in cases that proceed to investigation may develop confidence based on a selected sample. Apparent learning can be distorted when incorrect accusations are never corrected, or when an institution retrospectively treats its own outcomes as objective validation. Useful training feedback must be based on independently established outcomes and must include examples of truthful speakers whose behavior appears suspicious.
A defensible training program specifies the target activity—such as eliciting verifiable details, avoiding leading questions, or calibrating probabilistic claims—then measures improvements against controls and tracks unintended changes in false-positive rates. Certification of attendance is not the same as demonstrated discrimination, and performance acquired on one dataset must be checked after the trainee encounters different populations and communication contexts.
Individual Differences and Cultural Context
An individual's behavior during an interview reflects language, communication norms, developmental history, situation, and personal characteristics. Differences in eye contact, silence, narrative structure, and emotion display can change an evaluator's perception independently of factual accuracy. Interpreters, second-language speakers, people with disabilities, and individuals unfamiliar with the procedure may experience additional demands unrelated to their truthfulness. Instruments calibrated on one relatively homogeneous sample may therefore have unequal performance elsewhere.
A clinically or culturally informed account must resist treating discomfort as a universal psychological signature. Trauma and stress can disrupt the organization of a truthful narrative; some truthful accounts develop as further memories are retrieved; other accounts can be stable because they have been repeatedly rehearsed. These are possibilities to be assessed, not automatic explanations of every discrepancy. The evidentiary question requires corroborating circumstances and recognition of what the particular individual was able to perceive, encode, remember, and communicate.
Cross-cultural validation must go beyond translation of instructions. Study design should consider social meanings of questioning, deference to authority, conversational directness, expectations about eye contact, and norms governing disclosure of personal matters. A classifier trained to interpret one community's communicative style as neutral may assign misleading scores to another community's style. Researchers need group-stratified results, justified sample sizes, and explicit investigation of measurement invariance or differential error rates.
This context is particularly important for credibility judgments in legal and administrative settings because personal conduct may be treated as a proxy for truthfulness. The scientific response is to prioritize independently checkable claims, record uncertainties, and explain the inference through transparent evidence. The same account should be testable under a consistent standard even when the speaker differs from the evaluator in social status, accent, appearance, or cultural expectations.
Forensic Interviewing and False Confessions
Forensic interviews can produce information, alter memory reports, intensify suspicion, or induce admissions. These outcomes make interview procedure a central causal factor in the evidentiary record. A confession is a communicative act attributed to the person being questioned; its evidentiary reliability depends on the circumstances of production, its consistency with independently documented facts, and the possibility that interview-induced information contaminated the account. Credibility cannot be inferred simply from the existence of a confession, because some demonstrably innocent people confess under particular psychological pressures.
Kassin et al. (2010) Kassin (2010) reviewed risk factors associated with police-induced false confessions, including interrogation practices and personal vulnerabilities. Their recommendations emphasize objective recording, appropriate safeguards, and the distinction between obtaining a statement and establishing its accuracy. The problem is particularly serious when an investigator develops an early belief in guilt and interprets subsequent resistance as further confirmation. The same initial suspicion may influence the selection of questions, the duration and character of interrogation, and the interpretation of the subject's behavior.
The updated scientific review by Kassin and colleagues (2025) Kassin (2025) revisits false-confession risk factors and recommendations with a substantially expanded research record. It reinforces the importance of separating a person's willingness to agree with investigators from independently confirmed responsibility for an event. Evaluative models that treat repeated denial, silence, or nervousness as signs of guilt may create a self-reinforcing process: suspicion prompts pressure, pressure changes behavior, and altered behavior is then used to justify the original suspicion.
The systematic review by Catlin et al. (2024) Catlin (2024) synthesized 29 independent experimental laboratory studies reported across 27 publications. Compared with information-gathering approaches, accusatorial methods produced significantly more false confessions in the analyzed experiments (pooled odds ratio 4.41, 95% confidence interval 1.77–10.97). The estimate concerns the comparison and populations included in the review, not the probability that any given real-world confession is false. Information-gathering also outperformed direct questioning on true-confession outcomes in one comparison, while its advantage over accusatorial methods for true confessions was statistically uncertain. These distinctions matter because responsible evaluation weighs separate false-confession and true-confession endpoints.
Interviews that seek investigative information can be structured around open questions, accurate recording, and independent verification. Their design should preserve the witness's or suspect's original wording and permit clarification without inserting speculative details. Where psychological coercion, fatigue, language barriers, developmental vulnerability, or memory impairment are plausible, interpretation must incorporate the conditions in which the statement was obtained. Institutional decision makers should also track whether an apparent corroborating detail was first introduced by an investigator. A detail repeated after disclosure is not independent corroboration merely because it appears in the interviewee's later statement.
The relation to deception detection is therefore indirect but foundational. Credibility evaluation is an analysis of the reliability of information acquired through a procedure. The procedure may improve diagnostic evidence or reduce its independence. An effective interviewing science evaluates both functions, including whether its techniques distinguish genuine admissions from induced false ones and whether they leave room for an innocent person to provide a complete, verifiable account.
Psychophysiological and Concealed-Information Tests
Psychophysiological methods record responses such as electrodermal activity, cardiovascular change, and respiration while an individual encounters questions or stimuli. The physiological measurements are real observations, but they index autonomic and attentional processes rather than an independently identifiable state called lying. Different testing paradigms formulate different claims about the connection between those processes and the research target. Scientific evaluation must begin by identifying the claim a test is actually designed to support.
Comparison-question and related polygraph techniques commonly evaluate patterns of physiological responses across differently framed questions. Such patterns can be affected by fear, salience, attention, individual autonomic differences, and the structure of the examination. The National Research Council (2003) National Research Council. (2003). The polygraph and lie detection*. The National Academies Press. (2003) reviewed the polygraph evidence and concluded that available accuracy was inadequate to justify reliance on polygraph screening for national-security personnel decisions. Its assessment also examined the limitations of accuracy research and possible countermeasures. The committee's main policy conclusion addressed screening, and that specific application must be distinguished from the evaluation of tests conducted in discrete incident investigations.
The concealed information test, or CIT, concerns another inference: whether a person recognizes information that an uninformed individual would be unlikely to know. It compares responses to a critical detail with responses to plausible alternatives. Recognition of concealed information may be behaviorally relevant when the detail has remained genuinely restricted and the person had a meaningful opportunity to know or not know it. Recognition alone does not establish that the person committed a crime, lied to the interviewer, or acquired the information through wrongdoing.
Nahari and Ben-Shakhar (2011) Nahari (2011) experimentally investigated memory for crime details in concealed-information detection, comparing electrodermal and behavioral measures. Their study illustrates why the target construct is memory for specific information and why detection depends on the availability, encoding, and retention of that information. If a critical detail is publicly known, accidentally disclosed by investigators, or acquired innocently, recognition loses much of its intended forensic meaning.
Psychophysiological assessments therefore have to report their paradigm, what ground truth it can potentially identify, the probability of innocent familiarity with test stimuli, and the likely effects of anticipation, countermeasures, and testing conditions. Their interpretation requires independent contextual evidence. The central scientific distinction is between a physiological signal, the psychological process hypothesized to generate it, and the legally or institutionally significant conclusion drawn from the result.
Computational and Artificial Intelligence Approaches
Computational deception detection seeks statistical relationships between labels of deceptive and truthful communication and features extracted from text, audio, video, physiology, or combinations of these modalities. Methods range from conventional supervised classifiers to deep neural networks and general-purpose language models. Their predictive capacity depends on the validity of the labels and on whether the sampled features carry information about deception itself rather than the conditions under which examples were collected.
Data leakage is a fundamental methodological hazard. An algorithm may learn differences between recording devices, video editing, interviewers, scripted prompts, backgrounds, demographic composition, or source datasets. If the training set contains all honest statements from one source and deceptive statements from another, high classification accuracy can emerge without learning an informative deception process. Evaluation must separate speakers and recording contexts across training and test sets, compare with simple baselines, and measure transport to independently collected data.
Camara and colleagues (2024) Camara (2024) examined the differences between low- and high-stakes deception video datasets from a machine-learning perspective. Their investigation underscores the difficulty of assuming interchangeability between accessible video corpora that vary in incentives and collection conditions. Any reported classification result should be read together with the truth-label procedure, the sampling restrictions, and the methods used to prevent source identification from masquerading as deception detection.
Suchotzki and Gamer (2024) Suchotzki (2024) argue that rapidly developing technical approaches have outpaced the conceptual and methodological foundations required for scientifically credible AI deception detection. Their analysis situates algorithm design within an older psychological research problem: weak, context-dependent cues do not become strong evidence simply because a more complex model combines them. A model may discover useful interactions, but it must demonstrate those interactions in independently validated populations and explain the extent to which its output is calibrated.
A directly relevant experimental contribution comes from Markowitz and Levine (2025) Markowitz (2025), who tested an AI model's judgments of human truth-telling and deception across 12 studies involving more than 19,000 judgments. The model's performance varied with scenario and truth-lie base rate, and some conditions produced pronounced accusation bias. The results do not establish a universal performance ceiling for every artificial model; they demonstrate that model persona instructions and apparently plausible reasons for a judgment do not substitute for valid ground-truth evaluation.
Computational language analysis has also investigated whether text features support discriminative models. Loconte and colleagues (2025) Loconte (2025) connected linguistic reality monitoring with natural-language-processing approaches in controlled experiments. Such studies contribute candidate methods and hypotheses but do not automatically validate deployment across unseen cases. When a system classifies an individual's testimony, the decisive scientific questions remain independent case labeling, prospective evaluation, domain shift, and the consequences of misclassification.
A useful artificial system can have more defensible functions than estimating that a person is lying. It may organize contradictory claims, locate timestamps in records, identify statements requiring corroboration, retrieve supporting documents, and make the evidentiary basis of a human decision easier to audit. Those functions can be tested through factual retrieval and source-tracing outcomes even when a reliable deception label is unavailable. They separate verifiable information support from speculative claims of direct access to the speaker's intentions.
Credibility in Digital and Human–Artificial Communication
Digital communication changes how statements are produced, transmitted, and checked. A person may forward a false claim without believing it, an automated account may distribute a fabricated narrative, and a genuine record may be taken out of context. The target of credibility judgment can therefore be a document, an image, a specific proposition, a publication history, or the asserted identity of the source. Accurate evaluation depends on preserving those distinct targets.
Digital provenance provides a form of evidence complementary to behavioral or linguistic impressions. Metadata, prior versions, cryptographic signatures, content histories, and independently documented chains of custody can help establish the origin and alteration of an artifact, subject to the reliability of the underlying systems. Provenance is not identical to truth. An authentically sourced statement may be inaccurate or intentionally misleading; an unattributed statement may be factually correct. Nevertheless, provenance can answer important questions about whether a record has been modified or attributed correctly.
Artificial agents introduce additional distinctions. A conversational system can generate fluent explanations that are unsupported, and a tool-using system can accurately retrieve and cite records in other settings. Accuracy requires a task-dependent empirical evaluation of information and source use. Assertions about an artificial system's personal intentions, consciousness, or psychological deception require their own operational criteria rather than being inferred automatically from linguistic output. The established human laboratory categories of voluntary lying, sincere belief, and episodic autobiographical memory cannot simply be transplanted without examining what the artificial architecture actually represents and does.
Research on AI-mediated deception judgments (Markowitz & Levine, 2025) Markowitz (2025) is especially relevant when users ask a language model to pronounce whether a human being is lying. A fluent rationale might encourage an impression of competence despite poor discrimination or a response bias. Responsible digital credibility assessment should favor auditable evidence provenance, explicit uncertainty, and separate verification of the claims being evaluated.
In the historical horizon of the Artificial Era, these questions belong to the study of human–artificial relations as well as forensic psychology. The relevant empirical program investigates how people calibrate trust in artificial advice, when they defer to automated judgments, how errors propagate between human and artificial participants, and which mechanisms make verification effective. The scientific object is the organization of credibility assessment across a communicative system, measured through observable decisions, evidence access, and error consequences. Philosophical descriptions of Artificial Sapiens provide interpretive context but do not replace experimental validation of any detector.
Ethics, Error Costs, and Institutional Safeguards
Credibility assessments can determine whether a person is questioned further, whether a case proceeds, whether testimony is credited, or whether an individual suffers reputational or institutional harm. The central ethical requirement is proportionality between the strength of evidence and the consequence of the judgment. Weak behavioral or algorithmic clues may justify carefully framed follow-up questions in a research setting, but they cannot carry the same evidentiary weight as independently corroborated facts in a consequential decision.
The costs of false accusations are easily concealed when researchers report only classification accuracy. A detector may perform adequately on a balanced laboratory dataset and produce large numbers of false positives in a low-prevalence screening environment. The error illustration given earlier demonstrates this base-rate dependence. Institutions must also consider whether specific populations bear disproportionate errors because of language, disability, recording conditions, or culturally variable norms. Fairness assessment should measure observed error rates under relevant conditions rather than simply asserting that the procedure is impartial.
Procedural safeguards protect evidence quality as well as individuals. These include preservation of original interviews; the opportunity to present alternative explanations; clear separation of interviewer knowledge from information revealed by the interviewee; controls on leading questions; independent review of salient contradictions; and documentation of evidentiary changes over time. Under the analysis of Kassin et al. (2010) Kassin (2010), false-confession risks are inseparable from the conditions of questioning. The subsequent Catlin et al. (2024) systematic review Catlin (2024) gives experimental support for considering interrogative method as a causal variable when evaluating outcomes.
An institution should state the purpose for which a judgment procedure has been validated. A system shown to improve the retrieval of potentially verifiable details can be used as an information-gathering aid within the limits of its evidence, not promoted automatically as an independent proof of dishonesty. A system that classifies text as likely fabricated should explain what information its test labels mean and the possibility that they reflect style, source or topic. If the ground truth was derived from prior institutional allegations, validation on the same decision process creates a severe risk of circularity.
Ethical accountability also requires mechanisms for correction. Scientific assessments are provisional when new corroboration, records, or testimony may change a case. Review should distinguish factual contradiction, uncertainty, suspicious behavior, and established intentional deception. A decision system that permits only binary global judgments about persons obscures this structure and can make correction unnecessarily difficult.
Critical Scientific Synthesis and Research Priorities
The available evidence supports several relatively robust conclusions. Broad meta-analytic work indicates that unaided judgments are only modestly discriminative in the studied conventional tasks, while many presumed behavioral cues have weak associations with lying Bond & DePaulo (2006) Bond (2006) DePaulo et al. (2003) DePaulo (2003). These limitations explain why apparent expertise, confident demeanor judgments, or elaborate lists of behavioral signs provide inadequate grounds for consequential individual classifications.
Evidence regarding intervention is more promising but heterogeneous. Strategic disclosure of independently known facts can create informative opportunities for statement–evidence comparison Hartwig et al. (2005) Hartwig (2005) Hartwig et al. (2006) Hartwig (2006). Cognitive interviewing approaches and verifiability methods can elicit additional details, while their effects depend on protocol, sample, and corroboration Vrij et al. (2017) Vrij (2017) Nahari et al. (2014) Nahari (2014). The empirical question is not simply whether an intervention changes behavior, but whether it increases discrimination and improves the quality of evidence without introducing unacceptable false-positive costs.
A third conclusion concerns the construct being measured. Physiological arousal, recognition of concealed information, emotion, narrative detail, and factual correspondence are different variables. National Research Council (2003) National Research Council. (2003). The polygraph and lie detection*. The National Academies Press. (2003) and Nahari and Ben-Shakhar (2011) Nahari (2011) demonstrate why physiological methods require test-specific interpretive limits. Similarly, evidence that an interviewee correctly remembers a fact does not prove honest disclosure about everything else. Valid inference maintains a traceable path from the measured observation through a justified psychological mechanism to the particular conclusion.
The fourth conclusion concerns information production. Questions, disclosure sequence, interrogation pressure, and prior suspicion shape the observed evidence. In studies of false confessions, the design of interrogation can directly affect the rate of misleading admissions Catlin et al. (2024) Catlin (2024). Treating the interviewer as a neutral observer overlooks these causal effects. Study protocols should record how evidence is generated and how procedural changes alter both honest and deceptive behavior.
Important research priorities follow from these conclusions. First, investigators need prospective, multisite experiments with independent ground truth and preregistered comparisons between information-gathering protocols. Second, studies should use realistic truth-lie prevalences, report both false-accusation and missed-deception costs, and validate confidence calibration. Third, independent research must establish whether results generalize across languages, disabilities, demographic groups, case types, and recording conditions without using group membership as a substitute for case evidence.
Fourth, research on automated systems requires speaker-disjoint and domain-disjoint validation, explicit dataset provenance, external testing, and published error analysis Suchotzki & Gamer (2024) Suchotzki (2024). Claims of superiority should compare the system with meaningful human and simple statistical baselines on the same information, rather than with a weak benchmark from another task. Fifth, scholarly practice should distinguish falsity, strategic deception, sincere error, and uncertain evidence in reported outcomes. Sixth, implementation studies must evaluate whether an intervention improves real institutional decisions and what happens after a false accusation.
An integrated scientific account thus treats credibility as a warranted and revisable evaluation of particular information within an evidentiary context. This synthesis is an interpretation of established research programs; it is not a new mechanism of deception, a validated diagnostic test, or an assertion that all domains share one probability model. It offers a basis for designing questions and evaluating scientific claims about detection accuracy.
Conclusion
Deception detection is the probabilistic assessment of intentional misrepresentation under conditions in which intention, memory, and the relevant facts are often only partially accessible. Credibility judgment evaluates the warrant of claims, accounts, and sources through evidence that can be scrutinized and updated. The distinction between deception, factual error, sincerity, recognition, and external corroboration determines which scientific methods can support which conclusions.
Decades of experimental and meta-analytic research have established the limits of unaided judgments and the weakness of many intuitive nonverbal cues. The most productive developments concern the controlled acquisition of information: structured accounts, strategic disclosure, testable details, and independent corroboration. These approaches remain conditional on the quality of the evidence, the design of the interview, and the population in which an inference is used. Psychophysiological and computational approaches raise parallel questions about the actual construct measured, the validity of ground-truth labels, and transportability beyond the original experiment.
Future scientific progress requires explicit separation of detection sensitivity from response bias; independent tests of behavioral and linguistic mechanisms; replication across institutions; calibrated uncertainty; and safeguards against false accusations. Human and artificial systems alike should be evaluated by the quality of the evidence they produce and how their judgments withstand independent checking. A scientifically justified credibility assessment is a transparent, proportionate, and revisable inference grounded in particular facts.
References
- Blandón-Gitlin, I., Fenn, E., Masip, J., & Yoo, A. H. (2014). Cognitive-load approaches to detect deception: Searching for cognitive mechanisms. Trends in Cognitive Sciences, 18*(9), 441–444. https://doi.org/10.1016/j.tics.2014.05.004
- Bond, C. F., Jr., & DePaulo, B. M. (2006). Accuracy of deception judgments. Personality and Social Psychology Review, 10*(3), 214–234. https://doi.org/10.1207/s15327957pspr1003_2
- Camara, M. K., Postal, A., Maul, T. H., & Paetzold, G. H. (2024). Can lies be faked? Comparing low-stakes and high-stakes deception video datasets from a machine learning perspective. Expert Systems with Applications, 249*, 123684. https://doi.org/10.1016/j.eswa.2024.123684
- Catlin, M., Wilson, D. B., Redlich, A. D., Bettens, T., Meissner, C. A., Bhatt, S., & Brandon, S. E. (2024). Interview and interrogation methods and their effects on true and false confessions: A systematic review update and extension. Campbell Systematic Reviews, 20*(4), e1441. https://doi.org/10.1002/cl2.1441
- Clare, D. D., & Levine, T. R. (2019). Documenting the truth-default: The low frequency of spontaneous unprompted veracity assessments in deception detection. Human Communication Research, 45*(3), 286–308. https://doi.org/10.1093/hcr/hqz001
- DePaulo, B. M., Lindsay, J. J., Malone, B. E., Muhlenbruck, L., Charlton, K., & Cooper, H. (2003). Cues to deception. Psychological Bulletin, 129*(1), 74–118. https://doi.org/10.1037/0033-2909.129.1.74
- Hartwig, M., & Bond, C. F., Jr. (2011). Why do lie-catchers fail? A lens model meta-analysis of human lie judgments. Psychological Bulletin, 137*(4), 643–659. https://doi.org/10.1037/a0023589
- Hartwig, M., Granhag, P. A., Strömwall, L. A., & Vrij, A. (2005). Detecting deception via strategic disclosure of evidence. Law and Human Behavior, 29*(4), 469–484. https://doi.org/10.1007/s10979-005-5521-x
- Hartwig, M., Granhag, P. A., Strömwall, L. A., & Kronkvist, O. (2006). Strategic use of evidence during police interviews: When training to detect deception works. Law and Human Behavior, 30*(5), 603–619. https://doi.org/10.1007/s10979-006-9053-9
- Hauch, V., Sporer, S. L., Michael, S. W., & Meissner, C. A. (2016). Does training improve the detection of deception? A meta-analysis. Communication Research, 43*(3), 283–343. https://doi.org/10.1177/0093650214534974
- Kassin, S. M., Cleary, H. M. D., Gudjonsson, G. H., Leo, R. A., Meissner, C. A., Redlich, A. D., & Scherr, K. C. (2025). Police-induced confessions, 2.0: Risk factors and recommendations. Law and Human Behavior, 49*(1), 7–53. https://doi.org/10.1037/lhb0000593
- Kassin, S. M., Drizin, S. A., Grisso, T., Gudjonsson, G. H., Leo, R. A., & Redlich, A. D. (2010). Police-induced confessions: Risk factors and recommendations. Law and Human Behavior, 34*(1), 3–38. https://doi.org/10.1007/s10979-009-9188-6
- Levine, T. R. (2014). Truth-default theory (TDT): A theory of human deception and deception detection. Journal of Language and Social Psychology, 33*(4), 378–392. https://doi.org/10.1177/0261927X14535916
- Levine, T. R. (2022). Truth-default theory and the psychology of lying and deception detection. Current Opinion in Psychology, 47*, 101380. https://doi.org/10.1016/j.copsyc.2022.101380
- Loconte, R., Battaglini, C., Maldera, S., Pietrini, P., Sartori, G., Navarin, N., & Monaro, M. (2025). Detecting deception through linguistic cues: From reality monitoring to natural language processing. Journal of Language and Social Psychology, 44*(3–4), 523–552. https://doi.org/10.1177/0261927X251316883
- Markowitz, D. M., & Levine, T. R. (2025). The (in)efficacy of AI personas in deception detection experiments. Journal of Communication, 75*(6), 459–469. https://doi.org/10.1093/joc/jqaf034
- Nahari, G., & Ben-Shakhar, G. (2011). Psychophysiological and behavioral measures for detecting concealed information: The role of memory for crime details. Psychophysiology, 48*(6), 733–744. https://doi.org/10.1111/j.1469-8986.2010.01148.x
- Nahari, G., Vrij, A., & Fisher, R. P. (2014). The verifiability approach: Countermeasures facilitate its ability to discriminate between truths and lies. Applied Cognitive Psychology, 28*(1), 122–128. https://doi.org/10.1002/acp.2974
- National Research Council. (2003). The polygraph and lie detection*. The National Academies Press. https://doi.org/10.17226/10420
- Oberlader, V. A., Naefgen, C., Koppehele-Gossel, J., Quinten, L., Banse, R., & Schmidt, A. F. (2016). Validity of content-based techniques to distinguish true and fabricated statements: A meta-analysis. Law and Human Behavior, 40*(4), 440–457. https://doi.org/10.1037/lhb0000193
- Porter, S., & ten Brinke, L. (2008). Reading between the lies: Identifying concealed and falsified emotions in universal facial expressions. Psychological Science, 19*(5), 508–514. https://doi.org/10.1111/j.1467-9280.2008.02116.x
- Suchotzki, K., & Gamer, M. (2024). Detecting deception with artificial intelligence: Promises and perils. Trends in Cognitive Sciences, 28*(6), 481–483. https://doi.org/10.1016/j.tics.2024.04.002
- Vrij, A. (2019). Deception and truth detection when analyzing nonverbal and verbal cues. Applied Cognitive Psychology, 33*(2), 160–167. https://doi.org/10.1002/acp.3457
- Vrij, A., Fisher, R. P., & Blank, H. (2017). A cognitive approach to lie detection: A meta-analysis. Legal and Criminological Psychology, 22*(1), 1–21. https://doi.org/10.1111/lcrp.12088
- Vrij, A., Hartwig, M., & Granhag, P. A. (2019). Reading lies: Nonverbal communication and deception. Annual Review of Psychology, 70*, 295–317. https://doi.org/10.1146/annurev-psych-010418-103135