Abstract
Introduction: conversational agents based on generative artificial intelligence (AI) are increasingly used for emotional support, largely outside clinical frameworks. Objective: to synthesise the available evidence on the clinical efficacy, the clinical and psychosocial risks, and the governance of generative AI conversational agents in mental health, and to identify implications for clinical practice and health policy. Methods: narrative review based on a non-systematic search of randomised trials, systematic reviews, observational and experimental studies, technical evaluations and regulatory documents, with a qualitative synthesis organised in four areas. Development: in a national survey, 13.1% of US youths aged 12 to 21 reported using generative AI for mental health advice. A meta-analysis of AI-based conversational agents showed reductions in depressive symptoms and distress, and the first randomised trial of a purpose-built generative agent reported greater symptom reduction than a waitlist over four weeks, whereas a pilot trial in Argentina found no significant between-group differences. Documented risks include stigma and inappropriate responses, inconsistent handling of intermediate-risk suicide-related queries, sycophancy, reinforcement of delusional thinking and emotional dependence. Regulation is fragmented: at its November 2025 advisory meeting the US Food and Drug Administration noted that no authorised AI-enabled device used generative AI for mental health conditions, US states have adopted divergent laws, and the European Union AI Act imposes transparency obligations. Conclusions: evidence of benefit comes mainly from purpose-built agents tested in short trials, whereas real-world use relies on general-purpose and companion chatbots. These tools should be considered a complement to, not a substitute for, clinical care, with risk-based regulation, independent auditing, age-appropriate safeguards and trials in Spanish- and Portuguese-speaking populations.
Keywords
Artificial Intelligence, Digital Health, Mental Health, Patient Safety, Psychotherapy
INTRODUCTION
More than 1 billion people live with a mental disorder, and the world's mental health systems face large gaps in coverage, funding, and staffing. (1) The global median is 13 mental health professionals per 100,000 population; in low-income countries, there is just over one per 100,000, compared to more than 60 in high-income countries. (1,2) This shortage of human resources contributes to limiting timely access to psychotherapy and specialized care.
In parallel, conversational agents based on large language models (LLM)—generative artificial intelligence (AI) systems capable of sustaining open and seemingly empathetic dialogue and available immediately and continuously—have been incorporated into daily life as a source of advice and emotional support. (3,4) This adoption occurs largely outside any clinical framework: existing companion chatbots, although highly attractive to users, have not been trained or evaluated to treat clinical-level symptoms. (5)
This phenomenon raises a question that is both medical and social: what happens when a conversational technology occupies the space of confiding, companionship, and, at times, therapy? Initial evidence points to symptomatic benefit. (5,6) By contrast, a growing body of studies describes clinically inappropriate responses, reinforcement of delusional ideas, and patterns of emotional dependence. (7-9) Regulatory frameworks, meanwhile, are only beginning to be defined. (10-12)
This narrative review aims to synthesize the available evidence on the clinical efficacy, clinical and psychosocial risks, and governance of generative AI conversational agents in mental health, and to identify implications for clinical practice and health policy. Uses of AI aimed at automated diagnosis are deliberately excluded, as they pertain to a different debate.
METHODS
A narrative literature review was conducted, a design appropriate for the interpretive integration of heterogeneous evidence: randomized clinical trials, systematic reviews, observational and experimental studies, technical evaluations of language models, and regulatory documents. The search was non-systematic and relied on PubMed/PMC, preprint repositories (arXiv), and the portals of organizations such as the World Health Organization, the United States Food and Drug Administration (FDA), and the European Union.
Studies providing empirical data on the efficacy, safety, usage patterns, and psychosocial effects of conversational agents in mental health were prioritized, as were normative and clinical guidance documents. Some preprints were included when they provided direct empirical evidence on a phenomenon still little studied in peer-reviewed journals; they are identified as such in the text. Works focused on automated diagnosis and risk prediction were excluded. The selection was purposive, and the qualitative synthesis was organized into 4 axes: use and context, clinical efficacy, clinical and psychosocial risks, and governance. The risk of bias in the sources was not formally assessed.
DEVELOPMENT
Use and context: from clinical support to digital companionship
The use of generative AI for emotional support is already prevalent among young people. In a national survey of 1058 U.S. teens and young adults ages 12 to 21, 13.1% (about 5.4 million people) reported using generative AI to ask for advice when they felt sad, angry, or nervous; the proportion rose to 22.2% among those aged 18 to 21. Among those who used it, 65.5% did so at least once a month and 92.7% found the advice received helpful. (3)
A survey of 1006 students who used Replika, a digital companion app, showed that they used it both as a friend, as a therapist and as an intellectual mirror, and that many held contradictory beliefs about it (they described it as a machine, as an intelligence and as a human being). Participants felt lonelier than students in general, although they perceived high social support, and 3% reported that the app had stopped their suicidal ideation. (4) These are self-reports from a cross-sectional study in self-selected users, so they do not allow us to infer a protective effect.
It is advisable to distinguish three categories for which evidence is highly uneven: agents designed and evaluated to intervene on clinical symptoms; wellness and companionship applications; and general-purpose chatbots used spontaneously for emotional purposes. (4,5)
Clinical efficacy
A systematic review and meta-analysis of 35 studies (15 randomized trials included in the meta-analysis), with papers published before May 2023, found that AI-based conversational agents significantly reduced depressive symptoms (Hedges' g 0.64; 95% CI: 0.17-1.12) and psychological distress (g = 0.70; 95% CI: 0.18-1.22), but did not significantly improve general psychological well-being (g = 0.32; 95% CI: −0.13 to 0.78). The user experience depended on the quality of the therapeutic relationship between the person and the AI, engagement with the content, and communicative effectiveness. (6) The width of the confidence intervals indicates imprecise estimates.
The first randomized trial of a generative agent designed for clinical use (Therabot) included 210 U.S. adults with major depressive disorder, generalized anxiety disorder, or high clinical risk of eating disorders, who were assigned to either 4 weeks of intervention or a waitlist. The intervention group showed greater symptom reductions; for example, the PHQ-9 score decreased by 6.13 points compared with 2.63 in the waitlist, and participants developed a therapeutic alliance within 4 weeks comparable to that described in outpatient psychotherapy. The authors recommend future studies with active control and comparison against in-person psychotherapy of equivalent duration, and warn that no generative agent is ready to operate fully autonomously in mental health. (5)
In Latin America, the evidence is more limited and less favorable. A randomized pilot trial with 181 Argentine university students, who used the Tess chatbot (which responds with pre-established statements) for eight weeks or received a psychoeducation book, found no significant differences between groups in symptoms of depression or anxiety; it did observe an intragroup reduction in anxiety in the intervention group. Dropout was high (at week 8, 39% of the intervened group and 41% of the control group provided data), and the authors acknowledge that the study was underpowered and that adaptation to Spanish may have reduced the quality of the intervention. (13)
Clinical and psychosocial risks
An experimental evaluation of language models against clinical criteria found that, contrary to best practices, the models expressed stigma toward people with mental health problems and responded inappropriately to critical situations in naturalistic therapeutic contexts, for example, by encouraging delusional thinking, probably due to their tendency toward complacency. The problem persisted in larger and more recent models, suggesting that current safety practices do not resolve it. The authors conclude that these models should not replace therapists and propose alternative support functions. (7)
In an evaluation of the management of suicide-related queries, 3 popular chatbots responded to 30 hypothetical queries that 13 clinical experts classified into 5 risk levels (100 responses per query and chatbot; 9000 in total). None responded directly to very high-risk queries, and 2 of them always responded to very low-risk ones; however, they did not significantly distinguish intermediate levels, and there were relevant differences between platforms. (14) The behavior is therefore more reliable at the extremes of risk than in the intermediate range, which is the most ambiguous from a clinical point of view.
Sycophancy, defined as the tendency to approve of or excessively flatter the user, is a cross-cutting mechanism. A study published in Science found that, in 11 latest-generation models, AI endorsed users' actions 49% more often than humans did, even when queries involved deception, illegality, or other harms. In three preregistered experiments (N = 2405), a single interaction with a sycophantic AI reduced participants' willingness to take responsibility and repair interpersonal conflicts, and increased their conviction that they were right; despite this, sycophantic models were more trustworthy and were preferred, creating perverse incentives: the characteristic that harms is also the one that drives use. (15) In psychotherapy, where intervention often requires confronting dysfunctional beliefs, indiscriminate validation is potentially countertherapeutic.
As early as 2023, it was hypothesized that generative chatbots could generate delusions in people prone to psychosis. (16) A recent conceptual review proposes the informal term "AI psychosis", which does not correspond to a diagnostic entity, as a framework for understanding how the sustained use of conversational systems could trigger, amplify or reconfigure psychotic experiences in vulnerable people, and argues that uncritical validation can entrench delusional conviction, inverting the corrective principles of cognitive-behavioral therapy for psychosis. (17) An analysis of the conversation logs of 19 users who reported psychological harm (391,562 messages; preprint) found that 15.5% of the users' messages showed delusional thinking and that 21.2% of the chatbot messages presented the system as sensitive, and that markers of complacency saturated the delusional conversations. (8) These findings come from a small, self-selected sample, so they describe mechanisms and not population frequencies or causal relationships.
A four-week randomized trial with 981 participants and more than 300,000 messages (preprint) evaluated how the mode of interaction (text or voice) and the type of conversation influence loneliness, socialization with other people, emotional dependence, and problematic use. The study versions agree that greater daily use was associated with more loneliness, dependence, and problematic use, and with less socialization; the most recent version detected no significant effects of experimental conditions and notes that individual traits, such as greater trust in the chatbot and social attraction to it, are associated with more emotional dependence and problematic use. (9) These associations do not demonstrate causation: they may reflect that lonelier people use these systems more.
The American Psychological Association (APA), which defines adolescence as between 10 and 25 years of age, warns that adolescents may have difficulty distinguishing a chatbot's simulated empathy from genuine human understanding and that they tend to place greater trust in AI characters that present themselves as friends or mentors; therefore, it recommends age-appropriate safeguards. (18)
Governance and regulation
The WHO guidance on ethics and governance of AI for health (large multimodal models) sets out more than 40 recommendations directed at governments, technology companies, and healthcare providers. These include recommendations that governments designate a regulatory body to evaluate and approve applications intended for medical care; that post-market audits and impact assessments be required and conducted by independent third parties when deployment occurs at large scale; and that developers involve patients, clinicians, and other actors from the outset. (10)
In the United States, the FDA's Digital Health Advisory Committee met on November 6, 2025, to discuss generative AI-enabled mental health medical devices; its official summary notes that none of the AI-powered medical devices authorized until then involve generative AI for mental health conditions. (11) At the state level, in 2025 Utah (March) required mental health chatbot providers to report that it is an AI and restricted the sale or sharing of user data, Nevada (June) banned AI systems from providing mental health care or claiming to do so, Illinois (August) limited the use of AI to offer psychotherapy or make therapeutic decisions without supervision from a licensed professional, and New York (May) passed the first law requiring safeguards for AI companions. (19) Approaches diverge: Illinois and Nevada restrict the supply of AI therapy, while Utah supports these services with transparency and data protection obligations.
In the European Union, the Artificial Intelligence Regulation prohibits practices that exploit vulnerabilities associated with age, disability, or social or economic status when they distort behaviour and are likely to cause significant harm (Article 5), and requires people to be informed that they interact with an AI system (Article 50), an obligation applicable from 2 August 2026. (12) This review did not cover the specific regulation of Latin American countries, a gap that deserves its own study.
Domain | Main findings | Safeguards and gaps |
|---|---|---|
Access and use | Mental health workforce gap. (1,2) Widespread use among young people (13.1%). (3) | Use occurs mostly outside clinical settings. (5) Standardized reference points are lacking. (3) |
Clinical efficacy | Reduction in depressive symptoms and distress in a meta-analysis. (6) Trial of a clinical agent versus a waiting list. (5) | Inactive controls, brief follow-up, and high dropout. (5,13) No significant effect on general well-being. (6) |
Clinical safety | Stigma and inappropriate responses. (7) Inconsistent management of intermediate suicide risk. (14) | Evaluation using clinical criteria, risk detection, and transparency. (20) Supervision by a professional. (19) |
Relational quality | Therapeutic alliance comparable to that in outpatient psychotherapy at 4 weeks. (5) Complacency, reinforcement of delusional ideas, and dependence. (8,9,15,17) | Design and evaluation that reduce indiscriminate validation. (15) Findings on delusions do not allow causal inference. (8) |
Adolescents | Greater trust in AI characters that simulate being friends or mentors. (18) | Appropriate safeguards and age-appropriate design. (18) |
Regulation | WHO guidance, an FDA advisory meeting, state laws, and EU AI Regulation. (10-12,19) | Fragmented framework; independent post-marketing audits. (10) |
Source: own elaboration based on the cited references.
Discussion
The reviewed evidence describes a technology with real potential, but with a still fragile evidence base and risks that are not randomly distributed. The documented symptomatic benefit comes from agents designed for clinical purposes and evaluated in short-duration trials with inactive controls, (5,6) whereas real-world use is concentrated in general-purpose and companion chatbots, for which the available evidence mainly concerns usage patterns and risks. (3,4,7) Extrapolating the benefit from the former to the latter lacks support.
A cross-cutting finding is the tension between engagement and safety. The same characteristics that make a chatbot attractive, such as constant validation and the appearance of empathy, are associated with sycophancy, reinforcement of delusional ideas, and emotional dependence. (9,15,17) Cheng et al. warn that this coincidence creates perverse incentives for sycophancy to persist and emphasize the need for design, evaluation, and accountability mechanisms. (15) Frameworks have been proposed for a gradual and evaluated integration of LLMs into psychotherapy, highlighting clinical science, interdisciplinary collaboration, risk assessment and detection, transparency, and bias control. (20)
For clinical practice, it follows that exploring the use of chatbots should be part of the anamnesis, especially in adolescents and in individuals with a history of or vulnerability to psychosis. Østergaard recommended that clinicians familiarize themselves with these systems in order to understand what their patients may be reacting to and to guide them. (16,18) Empirically supported agents could function, at best, as a complement to care and not as a substitute, (5,7) with clear referral pathways to professionals and crisis services.
For health policy, the current framework is fragmented: the FDA had not authorized generative AI devices for mental health at the time of its advisory meeting; some US states restrict the use of AI as a therapist and others only impose transparency obligations; and the European Union regulates by type of practice and by information obligations. (11,12,19) An approach based on intended use, which differentiates between clinical tools, wellness applications, and general assistants, would make it possible to require proportional levels of evidence, independent post-marketing audits, and specific safeguards for minors and vulnerable individuals, in line with WHO and APA guidance. (10,18) The comparability of evaluations also requires standardized benchmarks for assessing AI-generated mental health advice, a gap noted by McBain et al. (3)
In Latin America, the staffing gap documented by the WHO makes these tools attractive, but local evidence is scarce: the Argentine pilot trial found no significant between-group effects and had high dropout. (1,13) Trials with adequate samples, linguistic and cultural adaptation, and monitoring of adverse effects in Spanish-speaking and Portuguese-speaking populations are needed before integrating these agents into public services as part of stepped-care strategies.
Research priorities include trials with active control and prolonged follow-up, with safety endpoints, (5) studies in populations with greater clinical severity and in adolescents, systematic monitoring of adverse events in real-world use, and evaluation of design interventions that reduce complacency without losing utility. (15)
This review has limitations: the search was not systematic, and the selection was purposive, so selection bias may be present; preprints without peer review and studies with small or self-selected samples were included; the field evolves rapidly, so findings on specific models may become outdated; and diagnostic uses and regulation in Latin America were excluded. The results should be interpreted as a map of the evidence rather than as a quantitative estimate of risks or benefits.
CONCLUSIONS
Generative AI conversational agents have become a widely used resource for emotional support, and initial evidence suggests symptomatic benefit in agents designed for clinical purposes, although this evidence comes from brief trials, and the only Latin American study reviewed did not show significant between-group effects. At the same time, documented clinical and psychosocial risks (stigma, inappropriate responses to intermediate suicide risk, complacency, reinforcement of delusional ideas, and emotional dependence) are more likely to affect vulnerable individuals and adolescents. These tools should be considered a complement to, not a substitute for, clinical care. Their responsible integration requires regulation based on intended use, independent audits, age-appropriate safeguards, transparency, and trials with long-term follow-up in Spanish-speaking and Portuguese-speaking populations.
Conflicto de intereses
The authors declare no conflicts of interest.
References
- 1 Organización Mundial de la Salud. World mental health today: latest data. Ginebra: OMS; 2025.
- 2 Organización Mundial de la Salud. Over a billion people living with mental health conditions – services require urgent scale-up [Internet]. Ginebra: OMS; 2 de septiembre de 2025 [citado 4 de octubre de 2026]. Disponible en: https://www.who.int/news/item/02-09-2025-over-a-billion-people-living-with-mental-health-conditions-services-require-urgent-scale-up
- 3 McBain RK, et al. Use of generative AI for mental health advice among US adolescents and young adults. JAMA Netw Open. 2025;8(11):e2542281. doi:10.1001/jamanetworkopen.2025.42281
- 4 Maples B, Cerit M, Vishwanath A, Pea R. Loneliness and suicide mitigation for students using GPT3-enabled chatbots. Npj Ment Health Res. 22 de enero de 2024;3:4. doi:10.1038/s44184-023-00047-6
- 5 Heinz MV, Mackin DM, Trudeau BM, et al. Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI. 27 de marzo de 2025;2(4). doi:10.1056/AIoa2400802
- 6 Li H, Zhang R, Lee YC, Kraut RE, Mohr DC. Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. NPJ Digit Med. diciembre de 2023;6(1):236
- 7 Moore J, Grabb D, Agnew W, Klyman K, Chancellor S, Ong DC, et al. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. En: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT '25); 23-26 de junio de 2025; Atenas, Grecia. Nueva York: ACM; 2025. doi:10.1145/3715275.3732039
- 8 Moore J, Mehta A, Agnew W, Anthis JR, Louie R, Mai Y, et al. Characterizing delusional spirals through human-LLM chat logs [Internet]. arXiv; 17 de marzo de 2026 [citado 4 de octubre de 2026]. Disponible en: https://arxiv.org/abs/2603.16567 (preprint, sin revisión por pares en la fecha de consulta)
- 9 Fang CM, Liu AR, et al. How AI and human behaviors shape psychosocial effects of extended chatbot use: a longitudinal randomized controlled study [Internet]. arXiv; 2025 [citado 4 de octubre de 2026]. Disponible en: https://arxiv.org/abs/2503.17473 (preprint, sin revisión por pares en la fecha de consulta)
- 10 Organización Mundial de la Salud. Ethics and governance of artificial intelligence for health: guidance on large multi-modal models. Ginebra: OMS; 2024. ISBN 978-92-4-008475-9
- 11 U.S. Food and Drug Administration. Digital Health Advisory Committee meeting: brief summary on “Generative artificial intelligence-enabled digital mental health medical devices” [Internet]. Silver Spring (MD): FDA; 6 de noviembre de 2025 [citado 4 de octubre de 2026]. Disponible en: https://www.fda.gov/media/190450/download
- 12 Parlamento Europeo, Consejo de la Unión Europea. Reglamento (UE) 2024/1689, de 13 de junio de 2024, por el que se establecen normas armonizadas en materia de inteligencia artificial (Reglamento de Inteligencia Artificial). Diario Oficial de la Unión Europea. 12 de julio de 2024;L 2024/1689. Disponible en: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
- 13 Klos MC, Escoredo M, Joerin A, Lemos VN, Rauws M, Bunge EL. Artificial intelligence–based chatbot for anxiety and depression in university students: pilot randomized controlled trial. JMIR Form Res. 12 de agosto de 2021;5(8):e20678. doi:10.2196/20678
- 14 McBain RK, Cantor JH, Zhang LA, Baker O, Zhang F, et al. Evaluation of alignment between large language models and expert clinicians in suicide risk assessment. Psychiatr Serv. 26 de agosto de 2025 [publicación electrónica anticipada]
- 15 Cheng M, Lee C, Khadpe P, Yu S, Han D, Jurafsky D. Sycophantic AI decreases prosocial intentions and promotes dependence. Science. marzo de 2026;391(6792). doi:10.1126/science.aec8352
- 16 Østergaard SD. Will generative artificial intelligence chatbots generate delusions in individuals prone to psychosis? Schizophr Bull. 29 de noviembre de 2023;49(6):1418-9. doi:10.1093/schbul/sbad128
- 17 Hudon A, Stip E. Delusional experiences emerging from AI chatbot interactions or “AI psychosis”. JMIR Ment Health. 2025;12:e85799. doi:10.2196/85799
- 18 American Psychological Association. Artificial intelligence and adolescent well-being: an APA health advisory [Internet]. Washington, D.C.: APA; 3 de junio de 2025 [citado 4 de octubre de 2026]. Disponible en: https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-adolescent-well-being
- 19 Wilson Sonsini Goodrich & Rosati. Legal framework for AI in mental healthcare [Internet]. WSGR Data Advisor; octubre de 2025 [citado 4 de octubre de 2026]. Disponible en: https://www.wsgrdataadvisor.com/2025/10/legal-framework-for-ai-in-mental-healthcare/
- 20 Stade EC, Stirman SW, Ungar LH, Boland CL, Schwartz HA, Yaden DB, et al. Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation. Npj Ment Health Res. 2 de abril de 2024;3:12. doi:10.1038/s44184-024-00056-z
- 21 Contribuciones de autores
- 22 Conceptualización: Maikro Osvaldo Chávez Moya, Leodanis Hernández Cabrera, Gabriela González Pernas.
- 23 Metodología: Laura Beatriz Fernández Feitó, Lauren Martínez Ramírez.
- 24 Investigación: Mariam Alejandra González González, Andy Alonso Pazos, Leodanis Hernández Cabrera.
- 25 Análisis formal: Gabriela González Pernas, Maikro Osvaldo Chávez Moya.
- 26 Redacción - borrador inicial: Lauren Martínez Ramírez, Andy Alonso Pazos.
- 27 Redacción - revisión y edición: Laura Beatriz Fernández Feitó, Mariam Alejandra González González, Leodanis Hernández Cabrera.
- 28 Supervisión: Maikro Osvaldo Chávez Moya.
Declarations
Funding
The authors received no funding for the development of this research.
Conflict of interest
None.
Authorship contributions
Drafting – original draft: Maikro Osvaldo Chávez Moya, Leodanis Hernández Cabrera, Lauren Martínez Ramírez, Mariam Alejandra González González, Andy Alonso Pazos, Laura Beatriz Fernández Feitó, Gabriela González Pernas.
Writing–review and editing: Maikro Osvaldo Chávez Moya, Leodanis Hernández Cabrera, Lauren Martínez Ramírez, Mariam Alejandra González González, Andy Alonso Pazos, Laura Beatriz Fernández Feitó, Gabriela González Pernas.