Covert lie detection using keyboard dynamics

Detecting lies in investigative interviews through the analysis of response latencies and error rates to unexpected questions

In this study, we propose an approach to detect deception during investigative interviews by integrating response latency and error analysis with the unexpected question technique. Sixty participants were assigned to an honest (n = 30) or deceptive group (n = 30). The deceptive group was instructed to memorize the false biographical details of a fictitious identity. Throughout the interviews, participants were presented with a randomized sequence of control, expected, and unexpected open-ended questions about identity. Responses were audio recorded for detailed examination. Our findings indicate that deceptive participants showed markedly longer latencies and higher error rates when answering expected (requiring deception) and unexpected questions (for which premeditated deception was not possible). Longer response latencies were also observed in participants attempting deception when answering control questions (which necessitated truthful answers). Moreover, a within-subject analysis highlighted that responding to unexpected questions significantly impaired individuals' performance compared to answering control and expected questions. Leveraging machine-learning algorithms, our approach attained a classification accuracy of 98% in distinguishing deceptive and honest participants. Additionally, a classification analysis on single response levels was conducted. Our findings underscore the effectiveness of merging response latency metrics and error rates with unexpected questioning as a robust method for identity deception detection in investigative interviews. We also discuss significant implications for enhancing interview strategies.

Verbal lie detection using Large Language Models

Human accuracy in detecting deception with intuitive judgments has been proven to not go above the chance level. Therefore, several automatized verbal lie detection techniques employing Machine Learning and Transformer models have been developed to reach higher levels of accuracy. This study is the first to explore the performance of a Large Language Model, FLAN-T5 (small and base sizes), in a lie-detection classification task in three English-language datasets encompassing personal opinions, autobiographical memories, and future intentions. After performing stylometric analysis to describe linguistic differences in the three datasets, we tested the small- and base-sized FLAN-T5 in three Scenarios using 10-fold cross-validation: one with train and test set coming from the same single dataset, one with train set coming from two datasets and the test set coming from the third remaining dataset, one with train and test set coming from all the three datasets. We reached state-of-the-art results in Scenarios 1 and 3, outperforming previous benchmarks. The results revealed also that model performance depended on model size, with larger models exhibiting higher performance. Furthermore, stylometric analysis was performed to carry out explainability analysis, finding that linguistic features associated with the Cognitive Load framework may influence the model's predictions.

Building a second-opinion tool for classical polygraph

Classical polygraph screenings are routinely used by critical businesses such as banking, law enforcement agencies, and federal governments. A major concern of scientific communities is that screenings are prone to errors. However, screening errors are not only due to the method, but also due to human (polygraph examiner) error. Here we show application of machine learning (ML) to detect examiner errors. From an ML perspective, we trained an error detection model in the absence of labeled errors. From a practical perspective, we devised and tested successfully a second-opinion tool to find human errors in examiners' conclusions, thus reducing subjectivity of polygraph screenings. We report novel features that uplift the model's accuracy, and experimental results on whether people lie differently on different topics. We anticipate our results to be a step towards rethinking classical polygraph practices.

Sharing the Space With the "Victim" Can Increase Help Rates. A Study With Virtual Reality

A typical protocol for the psychological study of helping behavior features two core roles: a help seeker suffering from some personal or situational emergency (often called “victim”) and a potential helper. The setting of these studies is such that the victim and the helper often share the same space. We wondered whether this spatial arrangement might affect the help rate. Thus, we designed a simple study with virtual reality in which space sharing could be manipulated. The participant plays the role of a potential helper; the victim is a humanoid located inside the virtual building. When the request for help is issued, the participant can be either in the same spatial region as the victim (the virtual building) or outside it. The effect of space was tested in two kinds of emergencies: a mere request for help and a request for help during a fire. The analysis shows that, in both kinds of emergencies, the participants were more likely to help the victim when sharing the space with it. This study suggests controlling the spatial arrangement when investigating helping behavior. It also illustrates the expediency of virtual reality to further investigate the role of space on pro-social behavior during emergencies.

Assessing the impact of previous experience on lie effects through a transfer paradigm

Influential lines of research propose dual processes-based explanations to account for both the cognitive cost implied in lying and for that entailed in the resolution of the conflict posited by Simon tasks. The emergence and consistency of the Simon effect has been proved to be modulated by both practice effects and transfer effects. Although several studies provided evidence that the lying cognitive demand may vary as a function of practice, whether and how transfer effects could also play a role remains an open question. We addressed this question with one experiment in which participants completed a Differentiation of Deception Paradigm twice (baseline and test sessions). Crucially, between the baseline and the test sessions, participants performed a training session consisting in a spatial compatibility task with incompatible (condition 1) or compatible (condition 2) mapping, a non-spatial task (condition 3) and a no task one (condition 4). Results speak in favour of a modulation of individual performances by means of an immediate prior experience, and specifically with an incompatible spatial training.

The detection of faked identity using unexpected questions and choice reaction times

The identification of faked identities, especially within the Internet environment, still remains a challenging issue both for companies and researchers. Recently, however, latency-based lie detection techniques have been developed to evaluate whether the respondent is the real owner of a certain identity. Among the paradigms applied to this purpose, the technique of asking unexpected questions has proved to be useful to differentiate liars from truth-tellers. The aim of the present study was to assess whether a choice reaction times (RT) paradigm, combined with the unexpected question technique, could efficiently detect identity liars. Results demonstrate that the most informative feature in distinguishing liars from truth-tellers is the Inverse Efficiency Score (IES, an index that combines speed and accuracy) to unexpected questions. Moreover, to focus on the predictive power of the technique, machine-learning models were trained and tested, obtaining an out-of-sample classification accuracy of 90%. Overall, these findings indicate that it is possible to detect liars declaring faked identities by asking unexpected questions and measuring RTs and errors, with an accuracy comparable to that of well-established latency-based techniques, such as mouse and keystroke dynamics recording.

Knowledge of Depression and Malingering: An Exploratory Investigation

Malingering mental disorder for financial compensation can offer substantial rewards to those willing to do so. A recent review of UK medico-legal experts' practices for detecting claimants evidenced that they are not well equipped to detect those that do. This is not surprising, considering that very little is known regarding why individuals opt to malinger. A potential construct which may influence an individual's choice to malinger is their knowledge of the disorder, and when one considers the high levels of depression literacy within the UK, it is imperative that this hypothesis is investigated. A brief depression knowledge scale was devised and administered to undergraduate students (N = 155) alongside a series of questions exploring how likely participants were to malinger in both workplace stress and claiming for benefit vignettes. Depression knowledge did not affect the likelihood of engaging in any malingering strategy in either the workplace stress vignettes or the benefit claimant vignettes. Differences were found between the two vignettes providing evidence for the context-specific nature of malingering, and an individual's previous mental disorder was also influential.

Machine Learning in Psychometrics and Psychological Research

Recent controversies about the level of replicability of behavioral research analyzed using statistical inference have cast interest in developing more efficient techniques for analyzing the results of psychological experiments. Here we claim that complementing the analytical workflow of psychological experiments with Machine Learning-based analysis will both maximize accuracy and minimize replicability issues. As compared to statistical inference, ML analysis of experimental data is model agnostic and primarily focused on prediction rather than inference. We also highlight some potential pitfalls resulting from adoption of Machine Learning based experiment analysis. If not properly used it can lead to over-optimistic accuracy estimates similarly observed using statistical inference. Remedies to such pitfalls are also presented such and building model based on cross validation and the use of ensemble models. ML models are typically regarded as black boxes and we will discuss strategies aimed at rendering more transparent the predictions.

Introducing Machine Learning to Detect Personality Faking-Good in a Male Sample: A New Model Based on Minnesota Multiphasic Personality Inventory-2 Restructured Form Scales and Reaction Times

Background and Purpose. The use of machine learning (ML) models in the detection of malingering has yielded encouraging results, showing promising accuracy levels. We investigated the possible application of this methodology when trained on behavioral features, such as response time (RT) and time pressure, to identify faking behavior in self-report personality questionnaires. To do so, we reintroduced the article of Roma et al. (2018), which highlighted that RTs and time pressure are useful variables in the detection of faking; we then extended the number of participants and applied an ML analysis. Materials and Methods. The sample was composed of 175 subjects, of whom all were graduates (having completed at least 17 years of instruction), male, and Caucasian. Subjects were randomly assigned to four groups: honest speeded, faking-good speeded, honest unspeeded, and faking-good unspeeded. A software version of the Minnesota Multiphasic Personality Inventory-2 Restructured Form (MMPI-2-RF) was administered. Results. Results indicated that ML algorithms reached very high accuracies (around 95%) in detecting malingerers when subjects are instructed to respond under time pressure. The classifiers' performance was lower when the subjects responded with no time restriction to the MMPI-2-RF items, with accuracies ranging from 75% to 85%. Further analysis demonstrated that T-scores of validity scales are ineffective to detect fakers when participants were not under temporal pressure (accuracies 55-65%), whereas temporal features resulted to be more useful (accuracies 70-75%). By contrast, temporal features and T-scores of validity scales are equally effective in detecting fakers when subjects are under time pressure (accuracies higher than 90%). Discussion. To conclude, results demonstrated that ML techniques are extremely valuable and reach high performance in detecting fakers in self-report personality questionnaires over more the traditional psychometric techniques. Validity scales MMPI-2-RF manual criteria are very poor in identifying under-reported profiles. Moreover, temporal measures are useful tools in distinguishing honest from dishonest responders, especially in a no time pressure condition. Indeed, time pressure brings out malingerers in clearer way than does no time pressure condition.

Unanticipated questions can yield unanticipated outcomes in investigative interviews

Asking unanticipated questions in investigative interviews can elicit differences in the verbal behaviour of truth-tellers and liars: When faced with unanticipated questions, liars give less detailed and consistent responses than truth-tellers. Do such differences in verbal behaviour lead to an improvement in the accuracy of interviewers’ veracity judgements? Two empirical studies evaluated the efficacy of the unanticipated questions technique. Experiment 1 compared two types of unanticipated questions (questions regarding the planning of a task and questions regarding the specific spatial and temporal details associated with the task), assessing the veracity judgements of interviewers and verbal content of interviewees’ responses. Experiment 2 assessed veracity judgements of independent observers. Overall, the results provide little support for the technique. For interviewers, unanticipated questions failed to improve veracity judgement accuracy above chance. Reality monitoring analysis revealed qualitatively distinct information in the responses to the two unanticipated question types, though little distinction between the responses of truth-tellers and liars. Accuracy for observers was greater when judging transcripts of unanticipated questions, and this effect was stronger for spatial and temporal questions than planning questions. The benefits of unanticipated questioning appear limited to post-interview situations. Furthermore, the type of unanticipated question affects both the type of information gathered and the ability to detect deceit.

False Identity Detection Using Complex Sentences

The use of faked identities is a current issue for both physical and online security. In this paper, we test the differences between subjects who report their true identity and the ones who give fake identity responding to control, simple, and complex questions. Asking complex questions is a new procedure for increasing liars' cognitive load, which is presented in this paper for the first time. The experiment consisted in an identity verification task, during which response time and errors were collected. Twenty participants were instructed to lie about their identity, whereas the other 20 were asked to respond truthfully. Different machine learning (ML) models were trained, reaching an accuracy level around 90–95% in distinguishing liars from truth tellers based on error rate and response time. Then, to evaluate the generalization and replicability of these models, a new sample of 10 participants were tested and classified, obtaining an accuracy between 80 and 90%. In short, results indicate that liars may be efficiently distinguished from truth tellers on the basis of their response times and errors to complex questions, with an adequate generalization accuracy of the classification models.

The Detection of Malingering: A New Tool to Identify Made-Up Depression

Major depression is a high-prevalence mental disease with major socio-economic impact, for both the direct and the indirect costs. Major depression symptoms can be faked or exaggerated in order to obtain economic compensation from insurance companies. Critically, depression is potentially easily malingered, as the symptoms that characterize this psychiatric disorder are not difficult to emulate. Although some tools to assess malingering of psychiatric conditions are already available, they are principally based on self-reporting and are thus easily faked. In this paper, we propose a new method to automatically detect the simulation of depression, which is based on the analysis of mouse movements while the patient is engaged in a double-choice computerized task, responding to simple and complex questions about depressive symptoms. This tool clearly has a key advantage over the other tools: the kinematic movement is not consciously controllable by the subjects, and thus it is almost impossible to deceive. Two groups of subjects were recruited for the study. The first one, which was used to train different machine-learning algorithms, comprises 60 subjects (20 depressed patients and 40 healthy volunteers); the second one, which was used to test the machine-learning models, comprises 27 subjects (9 depressed patients and 18 healthy volunteers). In both groups, the healthy volunteers were randomly assigned to the liars and truth-tellers group. Machine-learning models were trained on mouse dynamics features, which were collected during the subject response, and on the number of symptoms reported by participants. Statistical results demonstrated that individuals that malingered depression reported a higher number of depressive and non-depressive symptoms than depressed participants, whereas individuals suffering from depression took more time to perform the mouse-based tasks compared to both truth-tellers and liars. Machine-learning models reached a classification accuracy up to 96% in distinguishing liars from depressed patients and truth-tellers. Despite this, the data are not conclusive, as the accuracy of the algorithm has not been compared with the accuracy of the clinicians; this study presents a possible useful method that is worth further investigation.