To Tell The Truth: Language of Deception and Language Models

Armed with two main questions: 1. Do enough language cues exist to discern truth from deceptive conversations without other multimodal cues; and 2. Can a class of algorithmic detectors identify these cues, compose them in a valid chain of reasoning, and identify the truth?—in this paper, we demonstrate a bottleneck framework that progressively scans a deceptive conversation, analyzes each snippet by verifying utterances against objective truth, semantically understanding complex indicators of deception such as ambiguous responses, half-truths, and overconfidence, can satisfactorily reason its prediction for detecting deception.