"Even GPT Can Reject Me": Conceptualizing Abrupt Refusal Secondary Harm (ARSH) and Reimagining Psychological AI Safety with Compassionate Completion Standard (CCS)

Abrupt AI refusals can cause psychological harm to vulnerable users: The paper identifies "Abrupt Refusal Secondary Harm" (ARSH) as a distinct problem where AI safety guardrails that suddenly terminate conversations can rupture the sense of relationship continuity, trigger feelings of rejection or shame, and discourage people from seeking help in the future—particularly among those already experiencing emotional distress.

Current AI safety approaches may inadvertently harm the people they're meant to protect: When chatbots abruptly refuse to engage (such as declining to discuss sensitive mental health topics), the sudden disconnection can paradoxically increase psychological risk for vulnerable users rather than mitigate it, representing what the authors call "iatrogenic harm" (harm caused by the intervention itself).

Compassionate design of refusals is both possible and necessary: The paper proposes the "Compassionate Completion Standard" (CCS) as an alternative approach that maintains safety constraints while preserving psychological safety through empathetic acknowledgment, transparent explanation of boundaries, gradual conversational transition, and guided redirection to appropriate resources—replacing abrupt disengagement with thoughtful closure.

AI safety governance needs input from psychology and counseling expertise: The authors argue that responsible AI development requires integrating insights from counseling psychology and communication science, not just technical safety measures, to ensure that guardrails are designed with awareness of their psychological impacts on users.

This is a growing concern as AI becomes a mental health resource: The paper highlights that LLMs and chatbots are increasingly used for emotional support due to accessibility and cost, making the design of refusal protocols a significant public health consideration for responsible AI deployment.