Abstract
Background: The intersection of a global psychiatric workforce shortage and the rapid advancement of Generative Artificial Intelligence (GenAI) has catalyzed the integration of Large Language Models (LLMs) into digital healthcare. LLMs are increasingly deployed as preliminary triage agents to assess psychological distress, yet their diagnostic reliability, demographic neutrality, and safety protocols in crisis scenarios remain heavily debated within the psychiatric community. Objective: This review evaluates current research to determine the trustworthiness of large language models (LLMs) used in psychiatric screening. In doing so, it shines a light on the algorithmic biases present in practical use and maps out the necessary ethical standards to ensure digital health services remain safe and fair for all patients. Methodology: We sourced peer-reviewed publications released between 2023 and 2026 by running a systematic literature search across Google Scholar, PubMed/MEDLINE, and IEEE Xplore. The synthesis focused on comparative accuracy studies, observational evaluations, and sociotechnical analyses of leading LLMs (e.g., GPT-4, Claude 3.5 Sonnet, Gemini 1.5 Pro) applied specifically to psychiatric screening and automated clinical routing. Results: LLMs demonstrate high diagnostic concordance with human clinicians in structured assessments for high-prevalence conditions like major depressive disorder, effectively parsing unstructured clinical narratives. However, they exhibit a consistent tendency toward conservative over-triage in psychiatric emergencies. Profound algorithmic biases persist; models frequently recommend divergent, often inferior, treatment pathways for marginalized demographic groups and display gender-based stigmatization. Furthermore, models struggle with complex psychopathologies and display inconsistent safety mechanisms during acute suicidal ideation queries. Conclusion: While LLMs offer substantial utility for scaling preliminary psychiatric screening and reducing administrative triage burdens, their propensity for factual hallucination, demographic bias, and uncalibrated risk escalation currently precludes autonomous clinical deployment. Robust, human-in-the-loop governance and culturally representative training corpora are imperative prior to further integration into global digital health ecosystems.

This publication has 25 references indexed in Scilit: