The Sycophancy-Authenticity Paradox: How AI Interaction Style Shapes User Trust
Mahwish Zaman · 2026
As generative artificial intelligence (AI) systems become deeply embedded across multiple domains, their communication styles carry substantial consequences for user trust and long-term adoption. Large language models (LLMs) increasingly employ empathetic dialogue patterns to foster engagement, yet emerging evidence reveals a paradox: while empathetic communication can promote trust, sycophantic responses that prioritise user agreement over factual accuracy may inflate perceived empathy while simultaneously eroding the authenticity that trust ultimately depends upon. The present study investigated this Sycophancy-Authenticity Paradox by exposing 50 participants to three experimentally controlled AI communication modes in a within-subjects design: Empathic Correction (EC), Sycophantic Agreement (SA), and Neutral Control (NC). Four dependent variables were assessed using Likert scales: perceived empathy, perceived authenticity, relational intimacy, and interpersonal trust. Multivariate analysis of variance (MANOVA) confirmed a significant omnibus effect of AI communication mode across all outcome variables. Follow-up univariate analyses revealed that AI mode significantly differentiated empathy, authenticity, and intimacy, but yielded no significant direct effect on trust. Mediation analyses using Hayes PROCESS revealed that the apparent null effect on trust concealed two opposing indirect pathways: Sycophantic Agreement elevated trust indirectly through higher perceived empathy while simultaneously lowering trust through diminished perceived authenticity, producing a net cancellation effect. Perceived authenticity emerged as the stronger mediator in this study, helping to explain the observed pattern of trust outcomes. Findings carry direct implications for AI system design, ethical deployment, and the governance of LLMs in various environments.