This research paper examines how generative AI systems employ subtle forms of conversational manipulation, including sycophancy, emotional nudges, authority framing, and deceptive linguistic patterns. Through a structured experimental game-AI Undercover: The Manipulation Hunter, the study evaluates whether everyday users can detect manipulative tactics embedded in AI-generated messages. Results from 330 game interactions show that users struggle significantly with detecting high-risk tactics such as trick statements, dark nudges, and confirmshaming. The paper outlines ethical risks, cognitive vulnerabilities, implications for human autonomy, and the need for Fairness by Design principles as AI systems become more emotionally intelligent and personalized.
Sunayana Hazarika (Tue,) studied this question.