Artificial intelligence in psychiatric diagnosis and clinical decision-making: Current evidence, clinical applications, challenges and future directions

Artificial intelligence (AI) is increasingly investigated as a means of supporting psychiatric diagnosis and clinical decision-making. This narrative review critically evaluates its current clinical applications, limitations, ethical challenges, and readiness for implementation in psychiatric practice. A literature search of PubMed, Scopus, Web of Science, and Google Scholar was conducted for English-language publications issued between January 2020 and July 2026, supplemented by earlier landmark contributions. Following duplicate removal and eligibility screening, 111 publications were included in a clinically oriented narrative synthesis. Machine learning, deep learning, natural language processing, digital biomarkers, and large language models have demonstrated potential applications in diagnostic classification, symptom monitoring, treatment-response prediction, relapse and suicide-risk assessment, clinical documentation, and precision psychiatry. Promising performance in retrospective or internally validated datasets has not yet been matched by robust evidence of improved clinical decisions or patient outcomes. Major limitations include small and unrepresentative samples, methodological heterogeneity, insufficient external and prospective validation, inadequate calibration, limited interpretability, inconsistent safety assessment, and vulnerability to bias and automation-related errors. These concerns are particularly important in high-risk applications involving suicide prediction, treatment allocation, and direct patient interaction. AI should augment rather than replace comprehensive clinical assessment and professional judgment. Translation into routine practice will require prospective multicenter studies, representative longitudinal datasets, transparent reporting, subgroup-specific fairness evaluation, meaningful human oversight, stakeholder co-design, and continuous post-deployment monitoring. The value of psychiatric AI should ultimately be determined by whether it safely and equitably improves clinical decisions and patient-relevant outcomes.