To Tell The Truth: Language of Deception and Language Models

Text-based false information permeates online discourses, yet evidence of people’s ability to discern truth from such deceptive textual content is scarce. We analyze a novel TV game show data where conversations in a high-stake environment between individuals with conflicting objectives result in lies. We investigate the manifestation of potentially verifiable language cues of deception in the presence of objective truth, a distinguishing feature absent in previous text-based deception datasets. We show that there exists a class of detectors (algorithms) that have similar truth detection performance compared to human subjects, even when the former accesses only the language cues while the latter engages in conversations with complete access to all potential sources of cues (language and audio-visual). Our model, built on a large language model, employs a bottleneck framework to learn discernible cues to determine truth, an act of reasoning in which human subjects often perform poorly, even with incentives. Our model detects novel but accurate language cues in many cases where humans failed to detect deception, opening up the possibility of humans collaborating with algorithms and ameliorating their ability to detect the truth.

Introduction. Deception is pervasive in conversational dialogues. Individuals motivated by self-interest often feel compelled to embellish the truth to promote their interests at the expense of others. Misleading communication, such as false testimony (Tetterton and Warren, 2005), fake news (Shu et al., 2017), identity fraud in dating sites (Lazarus et al., 2022), sock puppetry (Kumar et al., 2017), and propaganda campaigns (Allcott and Gentzkow, 2017), abundant daily, impacts political, social, and economic outcomes. This exchange of information leading to the decision of who and what to believe necessitates the tacit development of truth detection capability during conversations (Bond and DePaulo, 2006). In what follows, we explore if textual cues may increase the likelihood of fraud detection even in the presence of more overt visual or aural indicators. Consider the CEO scam, when fraudsters act as company executives to trick a victim into sending unauthorized wire transfers or divulging private information through email.

Discussion / Conclusion. In this paper, we first showed the existence of a class of algorithmic detectors based on LLMs that can successfully identify language cues of deception without the presence of other visual or audio cues. We contribute a novel dataset T4TEXT for deception detection in the presence of objective truth and achieve a model performance comparable to human performance. We further find that our best model performs well in cases where humans perform poorly and discover novel language models that could augment human reasoning to detect deception, opening up the possibility of human-LLM collaborations to combat misinformation. This paper advocates for human-AI collaboration, emphasizing the need for additional evidence on human dependence on algorithms in detecting textual deception.