Answers are grounded in this paper and cite their sources.
Objective Generative artificial intelligence (GenAI) chatbots increasingly function as an informal, de facto source of mental health support. This paper develops a governance framework for associated harms and specifies requirements for governance-ready design and evaluation.Method We conducted a purposive narrative review across seven databases (PubMed, PsycINFO, Scopus, Web of Science, ACM Digital Library, IEEE Xplore, Google Scholar; searches through July 2026) using structured search query strings. Empirical evidence was systematically categorised and distinguished from theoretical psychoanalytic constructs. Findings informed a foundational design analysis of Eva, an unvalidated research case architecture informed by lived experience input.Results We distinguish acute Type I harms (single-turn technical failures) from cumulative Type II harms (multi-session relational risks), characterised through digital psychological harm, the uncanny setting, relational hallucination, algorithmic collusion, AI-mediated psychosis, and epistemic exploitation. Governance-by-design responses include the SAFE-AI (i.e., Screening, Alignment, Facilitation, and Evaluation of AI-Enhanced Interventions) protocol, empathy by design, suitability screening, psychodynamic transcript analytics, bounded co-regulation, data sovereignty, and crisis-safe escalation rather than simulated therapy.Conclusions The central proposition of this paper is that generative AI may support mental health care only when bounded by governance-by-design, clinically informed safety mechanisms, and accountable human oversight, and only once these safeguards have been independently validated. Specifically, we identify three sets of conditions for governance-ready GenAI: clinical conditions, including suitability screening and crisis-safe escalation; relational conditions, including bounded roles, anti-collusive friction, and longitudinal monitoring of Type II drift; and operational conditions, including human oversight, data sovereignty, auditability, and independent validation.
No takes yet. Share an insight, caveat, or question.
Balcombe et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: