How Large Language Models Weight Narrative Cues When Detecting Deception in Oral Accounts: A Multi-Model Comparison

Abstract

Lie detection research has increasingly turned to artificial intelligence, yet the application of large language models (LLMs) remains underexplored for transcribed oral narratives. We compared five LLMs (Gemini-3-flash, Gemini-3-pro, Deepseek-v4-flash, GPT-5.6-luna, and Grok-4) in classifying truthful and deceptive statements from interview transcripts of oral travel narratives collected from 95 Chinese-speaking participants. Guided by Criteria-Based Content Analysis (CBCA), Reality Monitoring (RM), and the Verifiability Approach (VA), theory-driven prompts instructed models to score five narrative cues before making binary veracity judgments. Gemini-3-flash achieved the highest accuracy (74.74%, AUC = 0.76) with 81.63% sensitivity and 67.39% specificity, and stable repeated-run performance. For this optimal model, unusual/superfluous details served as the sole significant predictor of objective veracity even after controlling for text length. LLMs can detect deception above chance but show large performance heterogeneity. Independent human coding, cross-cultural replications, and adversarial robustness testing are needed before real-world deployment.