Beyond the Assigned Label: The Predictive Role of AI Acceptance and Disclosure Labels in Mental Health Chatbots

Abstract

As AI-driven chatbots increasingly support mental health interventions, understanding how their assigned disclosure label influences user experience is crucial. This mixed-methods study ( N  = 187; recruited via Prolific; ages 19–68; 52% female, 48% male - Given the sample’s concentration among Prolific users from South Africa and other English-speaking regions with relatively high digital literacy, transferability to broader clinical, cultural, and digital-access contexts should be interpreted cautiously).employs a randomized experimental design to test whether the assigned disclosure label (AI-Aware vs. Human-Unaware) and advice style (Empathetic vs. Rational), alongside users’ Theory of Change (ToC) and baseline AI acceptance, shape user perceived empathy, satisfaction, and session impact in a single-session mental health chatbot interaction. These null findings should be interpreted as the absence of statistically significant differences between assigned conditions in this specific design, rather than as evidence of equivalence or absence of psychological or relational effects. However, these null results must be interpreted with caution; the absence of explicit manipulation checks limits our ability to definitively conclude that the label successfully altered users’ internal beliefs about the agent’s true identity. In contrast, higher baseline AI acceptance robustly predicted greater perceived empathy and session impact, with an Acceptance × Disclosure interaction for empathy. Qualitatively, topic and sentiment patterns in conversation logs were highly similar across conditions. We argue that, in this brief, online, non-clinical, single-session context, baseline AI acceptance explained more variance in immediate user-reported outcomes than the assigned identity disclosure label. Practically, these findings suggest a hypothesis for future research: pre-intervention education designed to improve user acceptance may benefit deployment efforts, though this was not tested as a causal intervention here.