A narrative review of the accuracy and clinical readiness of artificial intelligence methods for detecting depression

Depression is a leading cause of disability worldwide, yet it remains frequently under-recognised because diagnosis relies on self-report and clinical interviews that are subjective, stigma-sensitive and resource-intensive, particularly in low-resource settings. This narrative review examines artificial intelligence (AI) approaches for detecting depression from text and clinical records (including large language models), speech and acoustic signals, facial behaviour, electroencephalography (EEG) and multimodal combinations. PubMed/MEDLINE, Embase, PsycINFO, Scopus, IEEE Xplore and the ACM Digital Library, supplemented by the ACL Anthology, arXiv and citation searching, were searched for studies published between January 2015 and August 2026, and studies were appraised for sample size, data modality, reference standard and validation strategy. Across modalities, many studies report accuracies above 80%, and multimodal fusion often outperforms single modalities. However, the highest figures (93–100%) come mainly from small case–control or benchmark datasets of 30–200 participants evaluated with internal validation, whereas studies with several hundred to almost 15,000 participants report substantially more modest performance (area under the curve or balanced accuracy of approximately 0.69–0.78 for language, speech and record-based models, and 54–61% accuracy for neuroimaging). Most models are trained against self-report scales such as the PHQ-8 rather than diagnostic interviews, external validation is rare, widely used benchmarks contain demographic and interviewer-prompt biases, and no study has yet shown that AI-based detection improves patient outcomes. Large language models extend text-based assessment but show variable, prompt-dependent performance, particularly when restricted to locally deployable models, and EEG offers objective biomarkers whose near-perfect accuracies in small samples have not been replicated at scale. AI therefore offers plausible added value for scalable pre-screening, continuous monitoring and triage where specialist capacity is limited, but it should currently be regarded as a clinician-supervised adjunct rather than a diagnostic replacement. Progress will require larger and more diverse datasets, interview-based reference standards, external and prospective validation, transparent reporting, fairness auditing and privacy-preserving design.