Artificial Intelligence in Mental Healthcare: A Critical Narrative Review of Diagnosis, Treatment Personalisation and Patient Monitoring

Abstract

Artificial intelligence has been proposed as a corrective to three persistent problems in mental healthcare: diagnostic imprecision, the trial-and-error character of treatment selection, and the episodic nature of clinical monitoring. The volume of primary research has expanded rapidly, yet few tools have altered routine practice. This critical narrative review examines evidence across the three domains in which artificial intelligence has been most extensively applied to mental health, namely diagnostic classification and risk detection, treatment personalisation, and continuous patient monitoring, and asks why demonstrated technical performance has so rarely converted into demonstrated clinical benefit. Literature was identified through a bibliographic metadata registry, a biomedical citation index, targeted searching of scholarly and institutional sources, and backward and forward citation tracking, covering January 2015 to 11 June 2026, with earlier work retained where conceptually necessary. Evidence was appraised for design adequacy, validation strategy, sample representativeness, outcome definition and reporting transparency, then synthesised thematically rather than study by study. Three findings recur. Apparent accuracy is systematically inflated by internal validation, small and selected samples, and reference standards of limited reliability; where external validation has been attempted, discrimination frequently falls towards chance. The three domains differ markedly in evidential maturity, since monitoring and conversational intervention now rest on randomised evidence and pooled effect estimates, whereas diagnostic classification and treatment-response prediction remain largely at the model-development stage. The binding constraints on translation are infrastructural and epistemic rather than algorithmic, encompassing narrow training populations, unreliable outcome labels, absent prospective evaluation and immature governance. Unresolved questions include whether any model confers benefit over routine care in prospective use, how algorithmic outputs should enter clinical judgement, and how safety should be established for generative systems operating outside professional supervision. Progress will depend less on model refinement than on representative longitudinal datasets, standardised outcome definitions, prospective impact evaluation and governance capable of distinguishing wellness products from clinical instruments.