JMIR Publications

ABSTRACT

Conversational artificial intelligence (AI) systems are increasingly used for emotional and mental health support because of their low cost, immediacy, and accessibility. However, when safety guardrails are triggered, these systems may abruptly refuse, redirect, or disengage from a conversation. In emotionally salient or high-risk contexts, the manner in which such interventions are delivered may intensify distress or contribute to additional harm among users who are already vulnerable. This viewpoint introduces abrupt refusal secondary harm (ARSH) as a preliminary construct describing potential psychological harm associated with abrupt, safety-driven changes in conversational engagement. Drawing on counseling and communication science as conceptual heuristics, we propose that abrupt refusal may disrupt perceived relational continuity, intensify emotional distress, and discourage future helpseeking. To address this possibility, we introduce the compassionate completion standard (CCS), a testable human-centered design hypothesis for safety interactions that aims to maintain appropriate boundaries while preserving relational coherence. The CCS emphasizes empathetic acknowledgement, transparent boundary setting, graded transition, and guided redirection rather than abrupt disengagement. Integrating awareness of ARSH into design practices may help identify avoidable secondary harms and inform AI safety practices that better account for users’ psychological wellbeing. Rather than presenting definitive empirical evidence, this viewpoint outlines an interdisciplinary research agenda and policy recommendations for the development and governance of conversational AI safety.