Abstract
Conversational systems based on large language models (LLMs) are being increasingly used as advisors in mental health and self-regulation contexts, yet causal evidence remains limited about whether such guidance strengthens human agency rather than shifting responsibility to the system. We propose a dual framework in which the advice style reflects two dimensions, namely a structural stance (power distance) and a relational stance (psychological safety). In an online vignette experiment in China (N = 980), participants sought job search guidance from an LLM and read either a baseline reply or one of eight discourse variants, while holding the advice content constant. Relative to the baseline, a low power distance and a high psychological safety increased the self-efficacy, whereas a high power distance and a low psychological safety decreased it. Combination conditions revealed an asymmetric constraint: when the power distance was high, the self-efficacy declined even when the psychological safety was high, suggesting that authority allocation can override relational reassurance. Mediation analyses showed that the perceived self-control accounted for 26.3% of the low power distance effect and perceived belongingness accounted for 40.9% of the high psychological safety effect, with no cross-mediation. Although mental health outcomes were not directly measured, our results position conversational stances as actionable levers that shape self-efficacy and agency-related mechanisms, which are critical for persistence and adherence in mental health-relevant behavior change.