When Flattery Backfires: How Sycophancy and Interaction Context Shape Perceived Authenticity and Trust in Large Language AI Models

Sycophancy in large language models (LLMs) is widespread and often exceeds typical human interactions. By validating users’ misconceptions, it risks fueling misinformation, eroding critical thinking, and undermining trust. Yet research rarely considers interactional context—when sycophancy enhances user perception versus when it backfires. We adopt a social sycophancy framework, defined as the excessive preservation of the user’s face, to examine contexts where queries carry implicit assumptions. A 2×2 online experiment (N=436) manipulated sycophancy level and context (support-seeking vs. advice-seeking). Results show that sycophancy reduces perceived authenticity, with a more pronounced decline in advice-seeking scenarios. Intriguingly, although sycophancy can directly enhance affective trust, it simultaneously undermines both affective and cognitive trust via the mediation of reduced perceived authenticity. We conclude with implications for more ethical, context-aware LLM design.