Would I Lie to You? Using Large Language Models for Deception Detection

Deception is prevalent in online interactions and poses a challenge to reliable information access. This paper explores the effectiveness of BERT and GPT in improving the accuracy of deception detection. Building on previous work with traditional machine learning models, we use BERT and GPT to analyze textual crowd-sourced statements on climate change and COVID-19. We use fine-tuned BERT models and employ zeroshot, few-shot, and Chain-of-Thought (CoT) techniques with GPT models to classify statements as truthful or deceptive. Our results show that these models significantly outperform traditional machine learning methods, achieving improved accuracy in distinguishing between truthful and deceptive statements. This success stems from the models' ability to capture contextual information and semantic nuances in the text. In addition, the flexibility of GPT in zero-shot and few-shot settings proves valuable for tasks with limited labeled data. These results highlight the potential of pre-trained language models in deception detection and pave the way for further advancements.