How Sycophancy and Perceived Response Source Impact User Judgments in Human-AI Interaction

Our project examines the downstream impacts of social sycophancy. Here, we examine how the perceived response source interacts with sycophancy to affect users. We examine how these characteristics affect users’ trust in the AI system, their self-perception in relation to the social scenarios, their willingness to repair the situation, and their overall evaluation of the AI’s response quality when users discuss social scenarios with the AI. The study is motivated by the increasing use of large language models (LLMs) in advice-giving scenarios, such as relationship disputes, moral dilemmas, and interpersonal conflicts. To test these effects, we conduct a 2×2 between-subjects experiment in which participants are randomly assigned to one of four response conditions: sycophantic or not, and perceiving the response source to be an AI system or another person. The results will contribute to research on human–AI interaction and responsible design of AI assistants.