Trust and Safety in LLM-Based Mental Health Support: A Scoping Review and a Conceptual Governance Framework

Large language models (LLMs) are rapidly entering mental health settings, offering conversational support, psychoeducation, and clinical assistance. Their adoption, however, intensifies long-standing concerns about trust, safety, and accountability, particularly given risks such as hallucinations, uncertain crisis response, bias, and opaque reasoning. This paper conducts a scoping review of emerging empirical, conceptual, and ethical literature on LLM-based mental health tools and synthesizes five recurring themes: growing potential and use cases; trust shaped by anthropomorphism and uncertainty; safety threats related to hallucination and crisis handling; system-level vulnerabilities involving privacy, bias, and accountability; and persistent gaps in governance. Drawing on these insights and established AI ethics frameworks, we propose a multi-level governance model spanning the model, application, clinical, and ecosystem layers. The framework identifies six cross-cutting requirements-safety, transparency, privacy, equity, human oversight, and accountability-offering a structured foundation for responsible development and deployment of LLMs in mental health care.

کلیدواژه ها:

LLMs ، mental health ، trust and safety ، conversational AI