GENERATIVE ARTIFICIAL INTELLIGENCE IN PSYCHIATRIC DOCUMENTATION: ACCURACY, ERRORS, AND HUMAN OVERSIGHT — A NARRATIVE REVIEW

Abstract

Introduction: Generative artificial intelligence can support psychiatric documentation, but fluent outputs may conceal clinical errors. Objective: To critically examine accuracy, documentation errors, and human oversight in the use of language models in psychiatry and mental health. Methods: A critical narrative review based on structured searches in PubMed/MEDLINE, LILACS via the Virtual Health Library, and SciELO, supplemented by Elicit, through September 2026; 15 studies formed the analytic core. Results: Studies reported potential gains in efficiency, organization, and readability, alongside unsupported content, omissions, distortions of meaning, and attribution errors. Performance varied substantially across the clinical documentation tasks, language models, and generation instructions examined. Human review by qualified professionals was frequently recommended, but its effectiveness in preventing errors was seldom directly evaluated. Discussion: Linguistic quality and clinician satisfaction do not establish clinical fidelity. Conclusion: Available findings favor assistive use with professional verification and institutional policies addressing safety, transparency, and governance; they do not support autonomous clinical documentation.