The proliferation of fake online reviews, a long-standing threat to platform trust, is now exacerbated by large language models (LLMs) capable of generating highly convincing deceptive text. Understanding the linguistic strategies LLMs employ is crucial for developing effective mitigation. To address this gap, we develop an explainable artificial intelligence (XAI)-based computational framework, grounded in deception detection theories, to analyze and distinguish the deceptive patterns of LLMs. A core component of our methodology is a novel Turing-style test designed for LLM-generated online reviews. When applied to three purpose-built datasets, our framework not only achieves high detection accuracy for both human-authored fakes (96.57 %) and LLM-generated fakes (96.13 %)—substantially outperforming two current general-purpose detectors—but also indicates that LLMs possess a human-level deceptive capability (metric gaps <0.72 %). The analysis reveals that while cues related to cognitive load and perceptual-contextual details are powerful discriminators for both human and machine deception, certainty uniquely signals LLM-generated text, whereas emotion is a primary predictor only for human fakes. These findings support a central dissociation hypothesis between linguistic generation and cognitive representation: LLM deception is characterized by strategies like surface-level fluency, content realism without experiential grounding, and positivity bias. This study probes the mechanistic differences between human and machine deception, delivers a robust computational detection framework, and advances the theoretical discourse on AI's capacity for deceit.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号