Verbal lie detection using Large Language Models
When producing deceptive narratives, liars employ verbal strategies to create false beliefs in the interacting partners and are thus involved in a specific and temporary psychological and emotional state1. For this reason, the Undeutsch hypothesis suggests that deceptive narratives differ in form and content from truthful narratives2. This topic has always been under constant investigation and development in the field of cognitive psychology, given its significant and promising applications in the forensic and legal setting3. Its potential pivotal role is in determining the honesty of witnesses and potential suspects during investigations and legal proceedings, impacting both the investigative information-gathering process and the final decision-making level4.
Decades of research have focused on identifying verbal cues for deception and developing effective methods to differentiate between truthful and deceptive narratives, with such verbal cues being, at best, subtle and typically resulting in both naive and expert individuals performing just above chance levels5,6. A potential explanation coming from social psychology for this unsatisfactory human performance is the intrinsic human inclination to the truth bias7, i.e., the cognitive heuristic of presumption of honesty, which makes people assume that an interaction partner is truthful unless they have reasons to believe otherwise8,9. However, it is worth mentioning that a more recent study challenged this solid result, finding that instructing participants to rely only on the best available cue, such as the detailedness of the story, enabled them to consistently discriminate lies from the truth with accuracy ranging from 59 to 79%10. This finding moves the debate on (1) the proper number of cues that judges should combine before providing their veracity judgment -with the suggestion that the use-the-best heuristic approach is the most straightforward and accurate- and thus on (2) the diagnosticity level of this cue.
More recently, the issue of verbal lie detection has also been tackled by employing computational techniques, such as stylometry. Stylometry refers to a set of methodologies and tools from computational linguistic and artificial intelligence that allow to conduct quantitative analysis of linguistic features within written texts to uncover distinctive patterns that can infer and characterize authorship or other stylistic attributes11– 13.
Alongside this trend, several recent studies have explored computational analysis of language in different domains, such as fake news16,17, transcriptions of court cases18– 20, evaluations of deceptive product reviews21– 23, investigations into cyber-crimes24, analysis of autobiographical information25, and assessments of deceptive intentions regarding future events26.
However, to the best of our knowledge, despite the extreme flexibility of LLMs, the procedure of fine-tuning an LLM on small corpora for a lie-detection task has remained unexplored.
Related works in the psychology field
Among previous psychological frameworks aimed at identifying reliable cues of verbal deception, the Distancing framework, the Cognitive Load (CL) theory, the Reality Monitoring (RM) framework, and the Verifiability Approach (VA) have been extensively studied, gaining empirical support for their efficacy not only from primary research but also from meta-analytic studies.
The Distancing framework of deception states that liars tend to distance themselves from their narratives as a mechanism to handle the negative emotions experienced while lying by using fewer self-references (e.g., "I,""me") and employing more other-references (e.g., "he," "they")3,29.
The CL framework states that liars consume more cognitive resources while fabricating their fake responses, checking their congruency with other fabricated information, and maintaining credibility and consistency in front of the examiner30, resulting in shorter, less elaborate, and less complex statements. A meta-analysis31 found that approaches based on CL theories produce higher accuracy rates in detecting deception than standard approaches.
The RM framework bases its assumptions on the memory characteristics literature hypothesizing that truthful recollections are based on experienced events, while deceptive recollections are based on imagined events32. Therefore, RM derives its predictions about truthful narratives from sensory, spatial, and temporal information and from emotions and feelings experienced during the event. On the contrary, predictions about deceptions are drawn from the number of cognitive operations (e.g., thoughts and reasonings)33–35. The total RM scores appear to be diagnostic (d = 0.55) in the detection accuracy of truthfulness36,37 (see also38 for an extensive review of verbal lie-detection methods). More recently, the RM framework was investigated through concreteness in language39. In this study, one underlying and partially supported assumption was the truthful concreteness hypothesis, which suggests that truthful statements usually consist of concrete, specific, and contextually relevant details. In contrast, deceptive or false statements often include more abstract and less specific information, being more associated with the RM criterion of cognitive operations.
The VA in verbal lie detection suggests that truthful statements are more likely to be verifiable than false or deceptive statements, as liars avoid mentioning details that could be verified with independent evidence to conceal their deception40,41. Verifiable details may be represented by activities involving or witnessed by identified individuals, documented through video or photographic evidence, or leaving digital or physical traces (e.g., phone calls or receipts)40,41.
Notably, these frameworks offer detectable linguistic cues that can be readily identified using NLP techniques and have been extensively studied in this sense.