AI Chatbots for Teen Mental Health Show Promise but Evidence Remains Thin, Major Review Finds

Artificial intelligence chatbots, machine learning screening tools, and large language models are being rapidly promoted as solutions to the world’s growing adolescent mental health crisis, but a new systematic mapping of the evidence reveals a field racing ahead of its own science. A scoping review published in Frontiers in Child and Adolescent Psychiatry examined five years of research on AI-enabled and comparator digital interventions for young people aged 10 to 24, and its conclusions are sobering: the strongest controlled evidence still belongs to older, human-led approaches, while the most futuristic tools rest on some of the weakest study designs.

The review, conducted by researchers at Universiti Kebangsaan Malaysia, Universiti Teknikal Malaysia Melaka, Multimedia University, Al-Aqsa University, and Universiti Tun Hussein Onn Malaysia, followed the PRISMA-ScR reporting guidelines and Joanna Briggs Institute methodology. The team searched six databases, including PubMed, Scopus, ProQuest, SpringerLink, Web of Science, and ScienceDirect, supplemented by manual and forward-citation searching. The sweep initially captured 10,962 records, but after deduplication and multi-stage screening by two independent reviewers, only 19 studies published between 2020 and 2025 met the predefined inclusion criteria. That attrition rate, from nearly eleven thousand records down to nineteen, is itself a signal of how fragmented and diffuse this emerging literature remains.

To bring order to a heterogeneous field, the researchers applied a novel analytic device: a Human-AI Collaboration Role framework that classifies every intervention along a four-level continuum, from fully human-led digital delivery with no AI component, through AI-assisted systems that supplement human decision-making, to AI-primary tools that lead the interaction with a human escalation pathway available, and finally fully automated systems operating without human oversight. When the 19 studies were coded against this framework, a striking pattern emerged. The most rigorous randomized and controlled evidence clustered at the lowest level of the continuum, occupied by therapist-guided internet-delivered cognitive behavioral therapy, or iCBT, which relies entirely on human-authored content and clinician oversight. By contrast, studies of AI-primary tools were predominantly observational, feasibility, or quasi-experimental in design, with limited randomized evidence.

The demographic picture was equally narrow. Most studies focused on depression and anxiety among mid-to-late adolescents, and 68.4 percent of the research was conducted in high-income countries, with the United States alone contributing five of the 19 studies. Only four studies enrolled participants exclusively under 18 years, while nine included samples spanning both minors and adults without age-disaggregated outcomes. Critically, the review found that no fully randomized diagnostic chatbot trials have been conducted with participants under 18, meaning that claims about AI chatbots diagnosing or screening mental health conditions in minors currently rest on no randomized evidence at all. One randomized non-diagnostic chatbot trial, a body-image intervention delivered to Brazilian adolescents aged 13 to 18, does exist, showing that chatbot-delivered interventions can be tested rigorously in this age group even as diagnostic applications remain untested.

Equity reporting emerged as one of the review’s most troubling findings. While age was reported in all 19 studies, socioeconomic status appeared in only 10.5 percent, race and ethnicity in 21.1 percent, and family involvement in 21.1 percent. Cultural adaptation was documented in 42.1 percent of studies, but typically amounted to surface-level translation or localization rather than genuine participatory co-design. Only one included study, a chatbot intervention for rural-living LGBTQ+ youth, involved adolescents as co-creators of the technology. The authors warn that machine learning models trained predominantly on data from high-income, largely White or East-Asian cohorts may systematically underperform when applied to adolescents from underrepresented ethnic, socioeconomic, or linguistic backgrounds, and that speech-based phenotyping tools are particularly vulnerable to dataset shift across cultural and linguistic boundaries.

The technical landscape mapped by the review spans an impressive range. Predictive machine learning models have been used to forecast mental health problems in adolescence from registry data in Sweden. Speech-based phenotyping studies have analyzed word usage in spontaneous speech to predict depressive symptoms among youth at high risk for mood disorders. Large language model systems, including evaluations of ChatGPT-4 against validated questionnaires for screening anxiety and depression in college students, represent the newest frontier. Chatbots such as Vitalk, Elomia, and GPT-based screeners reflect growing interest in semi-automated and generative approaches. Yet the authors caution that several of these tools, including the Vitalk evaluation, were authored by study-affiliated developers of the commercial products and drew on general-population samples rather than clinically characterized adolescent groups.

Safeguarding for minors emerged as a central concern. Crisis detection mechanisms, such as identifying self-harm ideation, severe distress, or rapid symptom escalation, were inconsistently implemented across the reviewed systems and rarely standardized. While some tools incorporated predefined risk thresholds or keyword-based triggers, comprehensive escalation pathways were often underreported, and most systems relied on passive recommendations rather than enforceable referral mechanisms. The authors argue that fully autonomous decision-making carries elevated risk in high-stakes youth mental health contexts given the current evidence base, and that human-in-the-loop oversight for diagnostic interpretation, crisis management, and care planning represents the more cautious approach available today.

Generative AI introduces its own distinct hazards. The review highlights hallucination, the tendency of large language models to produce inaccurate or clinically inappropriate responses, as particularly dangerous in emotionally sensitive contexts involving adolescents. Over-trust poses a parallel risk, since young users may anthropomorphize conversational agents or attribute clinical authority to systems that lack it. Privacy leakage through prompt logging, model training artifacts, or unintended data retention threatens some of the most sensitive personal information a teenager could disclose, and the non-deterministic nature of generative models complicates reproducible evaluation.

Implementation challenges compound the scientific ones. Most AI-enabled interventions were evaluated as standalone tools rather than embedded components within stepped-care pathways or clinician-supervised frameworks, and several studies reported increased workflow burden for clinicians when AI outputs required manual verification or duplicate documentation. School-based deployment faces additional hurdles, including coordination with educational staff, alignment with safeguarding policies, and access to referral networks. The review also notes a striking terminological inconsistency across the field, with terms like AI chatbot, conversational agent, virtual assistant, and digital mental health tool used interchangeably without consistent operational definitions, complicating comparison across studies.

Importantly, the authors are explicit about what their review does not show. Because the included studies were heterogeneous and no formal risk-of-bias assessment was undertaken, the findings represent a descriptive map of the current evidence landscape rather than an assessment of clinical effectiveness, safety, or comparative superiority. The review also acknowledges that its restriction to English-language publications may have systematically excluded relevant evidence from non-Western and low- and middle-income contexts. Looking forward, the team calls for adequately powered adolescent-focused randomized trials with longitudinal follow-up, participatory co-design involving adolescents and families, culturally informed adaptations, transparent ethical governance frameworks proportional to the degree of AI autonomy, and hybrid systems that balance automation with human oversight to preserve empathy, trust, and sustained engagement. Until those trials arrive, the promise of AI for teen mental health remains exactly that: a promise the evidence has not yet kept.

Subject of Research: AI-enabled digital interventions for adolescent mental health, including chatbots, screening tools, and human-AI collaboration frameworks

Article Title: AI-enabled and comparator digital interventions for adolescent mental health (2020–2025): early detection, human–AI collaboration, and ethical governance—a PRISMA-ScR scoping review

Article References: Yaacob, N. M., Doheir, M., Algumaei, A., Al-Andoli, M. N., Alsayaydeh, J. A. J., Elzamly, A., & Basari, A. S. (2026). AI-enabled and comparator digital interventions for adolescent mental health (2020–2025): early detection, human–AI collaboration, and ethical governance—a PRISMA-ScR scoping review. Frontiers in Child and Adolescent Psychiatry, 5, Article 1905782. https://doi.org/10.3389/frcha.2026.1905782

Image Credits: AI Generated

DOI: 10.3389/frcha.2026.1905782

Keywords: adolescent mental health, AI chatbots, digital mental health, scoping review, PRISMA-ScR, machine learning screening, large language models, human-AI collaboration, ethical governance, cognitive behavioral therapy, health equity, safeguarding minors