Large Language Models in Psychiatry and Mental Health Care: A Systematic Review of Clinical Use Cases, Evidence Maturity, and Translational Risk (Preprint)

BACKGROUND Large language models (LLMs) are rapidly emerging as tools with potential applications across mental health care, yet their implications for psychiatric practice remain unclear. OBJECTIVE This systematic review explores and categorises LLM use cases in mental health care to inform clinicians, researchers, and policymakers about emerging applications, opportunities, evidence gaps, and safety challenges relevant to psychiatric assessment, risk management, clinical documentation, and patient-facing support. METHODS We searched EMBASE, MEDLINE, PsycINFO, PubMed, the ACL Anthology, the ACM Digital Library, arXiv, medRxiv, and bioRxiv from 2017 to June 2025. Empirical studies evaluating LLMs for mental health care tasks were included. The review was reported according to PRISMA 2020. Records were screened against prespecified eligibility criteria, and data were extracted using a standardised form. Use cases were categorised using an iterative, data-driven taxonomy developed from the included studies. Evidence maturity, methodological concern, and reporting of safety, governance, oversight, and implementation detail were assessed using review-specific frameworks and synthesised narratively. RESULTS We identified 90 studies evaluating LLM applications in psychiatry and mental health care. Use-case categories were not mutually exclusive, as some studies addressed more than one application. The most commonly evaluated applications were data analysis or information extraction, early detection, risk stratification, conversational agents, disease classification, and diagnostic support, with fewer studies addressing psychotherapy, patient education or psychoeducation, management or treatment, and policy-related applications. Most studies used retrospective, benchmark-based, simulated, or offline evaluations, with fewer studies assessing prospective clinical use, real-world workflow integration, patient outcomes, or service-level impact. Most evaluations measured technical performance rather than patient, service, or implementation outcomes. Safety and governance issues were inconsistently evaluated. Recurrent limitations included hallucinated or inaccurate outputs, bias and cultural insensitivity, limited transparency and explainability, data privacy risks, uncertain accountability, and insufficient evidence of safety in high-risk clinical contexts. CONCLUSIONS Current evidence suggests that LLMs have potential across psychiatry and mental health care, but the evidence base remains preliminary and concentrated in retrospective, benchmark-based, or simulated evaluations. At present, these technologies are best understood as tools for supervised clinical augmentation rather than autonomous psychiatric care. Future research should prioritise prospective, use-case-specific evaluations that measure real-world outcomes, safety, equity, governance, and implementation feasibility. CLINICALTRIAL None required INTERNATIONAL REGISTERED REPORT RR2-10.2196/preprints.91348