Generative AI Mental Health Chatbot Interventions: A Scoping Review of Safety and User Experience

Generative Artificial Intelligence (GenAI) mental health chatbots promise scalable, on-demand support, yet GenAI-related harms have raised concerns about user safety. To our knowledge, this scoping review represents the first systematic mapping of chatbot architecture, delivery modalities, user experience (UX) outcomes, and risk mitigation strategies in purpose-built chatbot interventions. A systematic search of seven databases identified 1899 articles, from which 21 studies across 11 countries were included. Most interventions were early-stage, cognitive behavioural therapy-based, and delivered through non-embodied chatbots. Target conditions included depression, anxiety, dementia, eating disorder and post-traumatic stress disorder. UX outcomes indicated moderate-to-high usability, therapeutic alliance, and user satisfaction, driven by convenience, personalization, and perceived empathy. However, engagement commonly declined, attributed to trust concerns and limited interactivity. Risk mitigation strategies incorporated technical controls, most frequently fine-tuning and prompt engineering, followed by risk classifiers and content filters, retrieval-augmented generation, and rules-based hybrid integrations. These safeguards were often implemented alongside human oversight, data security measures, and crisis referral pathways. Adverse event monitoring was rare, but some studies documented missed crisis cues and inaccurate content. Future research is needed to evaluate efficacy, standardize safety and UX outcome measurement, and develop layered safeguards to enable integration into stepped or blended care.