Peer-reviewed article
Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation.
Meita Dhamayanti, Susaldi Susaldi, Fanni Hanifa, Yova Agustini, Ova Emilia et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This scoping review examined 103 studies on artificial intelligence tools designed to support mental health in adolescents. The review found that most AI applications focus on predicting mental health problems rather than treating them, and very few have been tested to show they actually help young people feel better. The evidence shows major gaps: almost no studies tested safety in real-world use, and almost none measured whether symptoms actually improved. The authors call for future AI mental health tools to be designed with strong protections for privacy, fairness, and crisis safety.
Key points
- Of 95 primary studies reviewed, machine learning was used in 44% and risk prediction was the main function in 47%
- Only 1 study (1.1%) reported symptom-level effectiveness, despite 87.4% reporting predictive performance
- No primary study provided dedicated prospective AI-safety evaluation
- Implementation research focused on bias monitoring (29.5%), privacy (27.4%), and human oversight (20.0%)
- No study evaluated sustained routine-care scale-up or implementation cost as a principal outcome
Limitations
This is a scoping review that maps the field rather than synthesising outcomes of interventions; it cannot quantify the magnitude of effectiveness or safety gaps, only describe what has and has not been studied. The review is limited to English-language peer-reviewed sources and does not include grey literature or applications in deployment without published evaluation.
Abstract
ABSTRACT: OBJECTIVES: To map artificial intelligence (AI)-enabled mental health applications for adolescents and distinguish the available evidence on clinical functions, predictive or diagnostic performance, symptom-level effectiveness, safety, governance and implementation readiness. DESIGN: Scoping review conducted using Joanna Briggs Institute guidance and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) extension for Scoping Reviews. DATA SOURCES: PubMed/MEDLINE, Scopus, the Cochrane Library and ScienceDirect were searched from 1 January 2014 to 17 June 2026. ELIGIBILITY CRITERIA: English-language peer-reviewed sources addressing AI-enabled screening, prediction, assessment, monitoring, therapeutic support, care navigation, safeguarding or implementation for adolescents or school-aged youth were eligible. Adult-only studies, non-AI digital tools, protocols without results and unverifiable or promotional sources were excluded. CHARTING AND SYNTHESIS: Primary studies and evidence syntheses were analysed separately. 95 primary studies were mapped using mutually exclusive principal AI-modality and clinical-function categories; effectiveness, safety and implementation domains were non-mutually exclusive. RESULTS: The database searches identified 11 167 records, and 1 additional evidence synthesis was identified and assessed during revision; 103 sources were included, comprising 95 primary studies and 8 evidence syntheses. Structured-data machine learning was the dominant modality (42/95, 44.2%), and risk prediction or prognosis was the leading clinical function (45/95, 47.4%). Predictive or diagnostic performance was reported in 83 studies (87.4%), but symptom-level effectiveness in only 1 (1.1%). No primary study provided a dedicated prospective AI-safety evaluation. Implementation evidence focused on bias or equity monitoring (28/95, 29.5%), privacy or data governance (26/95, 27.4%), human oversight (19/95, 20.0%), acceptability (18/95, 18.9%) and feasibility (15/95, 15.8%). No study evaluated sustained routine-care scale-up or implementation cost as a principal outcome. CONCLUSIONS: Within the retrieved evidence base, adolescent mental-health AI research was concentrated in technical development, with limited prospective evidence of clinical benefit, safety and routine implementation. The field is dominated by prediction and internal validation, whereas symptom-level benefit, prospective safety, workflow integration, scale-up and sustainability remain largely untested. Future applications should be evaluated as supervised, human-in-the-loop systems with explicit safeguards for consent, privacy, fairness, accountability and crisis escalation.