Current Landscape and Future Directions for Mental Health Conversational Agents for Youth: Scoping Review

📜🔗LINK to the source

1. Analytical Summary

1.1 Context and Issues

The review covers 39 empirical studies published between 2011 and February 2024, of which 72% were after 2021, conducted in North America, Europe, Latin America, and the Asia-Pacific. It focuses on those aged 15–24 (UN definition of youth) in a context of marked increase in mental health issues, distress, suicidal thoughts, and non-utilisation of care due to stigma and access barriers. Chatbots most often target general well-being, depression, anxiety, and stress, with some devices addressing body image, substance use, smartphone addiction, HIV/STIs, cancer, or type 1 diabetes. The majority of interventions concern “general” youth rather than highly vulnerable youth, although a recent trend is towards at-risk populations (LGBTQ+, chronic illnesses, migrants, etc.). However, the studies are of small scale (often <100 young people), of short duration (<4 weeks), and focused on acceptability and feasibility rather than on lasting clinical effects or adverse effects.

1.2 Operational Contributions

The article provides a detailed mapping of research characteristics (phases, designs, durations, samples), design choices (the role of the chatbot as a “peer” or “coach”, friendly/empathetic tone, low personalisation) and technical choices (prototypes, rules, NLP, mainstream messaging). It details the strengths perceived by young people (24/7 accessibility, anonymity, feeling of being listened to, multimedia content) and the major weaknesses (limited content, repetitive, poorly adapted, poor understanding of natural language, robotic responses, concerns about privacy). The discussion highlights security issues (suicidal risk, over-screening, substitution for human connection), long-term clinical relevance, inclusion (multiple disorders, diversity of audiences) and still non-existent ethical standards. Finally, guidelines are proposed to design agents that are truly youth-centred: co-design with young people and clinicians, scientifically based content, multimodality, personalisation, transparency about limitations, robust safety measures and integration with existing human resources.

2. Key points of the document

1. Research still immature and fragmented

   Most studies involve fewer than 100 young people, use user testing or short-term trial designs (<4 weeks) and are rather in the development and summative evaluation phase of prototypes, with little robust data on long-term clinical effectiveness. (p.4–7, Table 1, p.14–18)

2. Partial targeting of needs and audiences

   Chatbots primarily target well-being, depression, anxiety, and stress, with a recent increase towards at-risk populations (LGBTQ+, chronic illnesses, refugees), but they remain underrepresented on other common issues (ADHD, eating disorders, suicidality, self-harm) and are poorly suited for young people in significant vulnerability. (p.7–10, 17)

3. “Peer/coach” centred design, low personalisation, inadequate security

   Agents are often designed as “older peers” or friendly and empathetic coaches, sometimes culturally adapted, but personalisation options are rare (avatars, identities, content), and more than half of the tools lack real security functions (alerts, emergency contacts, reminders of non-substitution for care). (p.8–10)

4. Basic technical architecture and pre-written content

   The majority of devices are text-based prototypes deployed via messaging platforms (Facebook, WhatsApp, WeChat) or mobile applications, combining a few NLP components for free text and numerous rules/trees to deliver pre-written CBT or positive psychology content, without resorting to large generative language models. (p.10–13)

5. Strengths and weaknesses: accessibility vs poor content and little ethical consideration

   Young people appreciate 24/7 accessibility, anonymity, a non-judgmental tone, and visual media, but criticise repetitive, unintelligent, sometimes inaccurate, poorly personalised, overly textual, and anxiety-inducing responses regarding confidentiality; 90% of studies do not discuss ethical or security issues in a structured manner. (p.14–17)

2. Action points for local stakeholders

1. Develop a local reading grid before adopting a chatbot

   Use the dimensions of the paper (target audience, objectives, content, entry/exit modalities, personalisation, security, ethics) to build an analysis grid prior to any partnership with a chatbot publisher, requiring written elements on evidence of effectiveness, emergency procedures, data management, and regulatory compliance. (p.10–13, 17–19)

2. Embed the use in a structured protocol articulated to the “offline”

   If a chatbot is tested in a school, a university, a hospital service, a local mission, or a youth reception, define a protocol: clear information to young people about the limits of the tool, duration of use, clinical supervision modalities, explicit links to human resources (consultations, CUMP, emergency numbers, youth structures). (p.6–7, 16–19)

3. Co-design with young people and professionals

   Organise co-design workshops involving young people, parents, educators, psychologists, doctors, digital mediators to validate content, tone, dialogue scenarios, risk thresholds, and modalities for redirecting to professionals, particularly for LGBTQ+ audiences, migrants, young people with disabilities or chronic illnesses. (p.5, 7–9, 18–19)

4. Integrate safety safeguards and screen time limits

   Require any device to integrate modules for detecting risky words or expressions (suicide, violence, self-harm), immediate redirection to emergency services or trusted adults, and follow-up/nudge functions to limit screen time and avoid dependency on the chatbot to the detriment of human relationships. (p.9–10, 16–18)

5. Systematically document adverse effects and access inequalities

   In any local project, provide a standardised collection of incidents (inappropriate responses, misunderstandings, confidentiality issues, worsening symptoms, non‑access to care) and access barriers (literacy level, language, digital equipment, sensory disabilities), to feed into quality processes, risk assessments and decisions to continue or stop. (p.14–18, 20)

4. Additional References 

🔍➕ For more information, see the articles referenced by "Health Practices" on the digital and AI theme ➡️🔗 https://pratiquesensante.odoo.com/2-6-intelligence-artificielle-numerique

• Interventions by conversational agents for mental health: real effects, limitations and success conditions for prevention and care teams.
• Encyclical letter of His Holiness Pope Leo XIV Magnifica Humanitas on the protection of the human person in the age of artificial intelligence
• We do not know what the impact of the ban on social networks will be on young people, but we do it anyway!
• AI and support for vulnerabilities
• The best interests of the child in the digital environment

5. FAQ for professionals and volunteers

1. Can we recommend a chatbot to young people as a “solution” for mental health?

   Data shows short-term improvements in distress in some trials, but the studies are short, small and focused on prototypes: the tool can be a source of information and support, never a substitute for human care. (p.6–7, 14–18)

2. Which young people have been primarily studied?

   Mainly school-aged adolescents and young adults, often without severe psychiatric disorders, with mild to moderate symptoms of depression, anxiety or stress; young people in very high vulnerability are still underrepresented. (p.6–9, 17)

3. What concrete benefits do young people report?

   They appreciate the 24/7 availability, ease of access, friendly and non-judgmental tone, sense of anonymity, the opportunity to discuss sensitive topics, and multimedia content, sometimes associated with a short-term decrease in depression/anxiety scores. (p.14–16)

4. What major limitations emerge from the studies?

   Limited or repetitive content, insufficient understanding of natural language, sometimes inaccurate, poorly personalised or robotic responses, text overload, concerns about confidentiality and data processing, risks of screen overuse and substitution for human connection. (p.14–17)

5. What elements should be checked before deploying a chatbot in an organisation?

   Scientific quality of content (CBT, validated programmes), evaluation results (trials, indicators), data management methods (hosting, anonymisation, sharing), security functions (risk detection, emergency contacts), clarity of messages explaining the tool's limitations to young people. (p.10–13, 17–19)

6. How to integrate the tool with the existing offer?

   The document recommends positioning chatbots as a building block of a continuum: information, psychoeducation, guided self-help, follow-up tools, which direct towards human resources (consultations, school nurses, associative devices, helplines) rather than supplanting them. (p.16–19)

7. What points not covered by the studies should raise concern?

   The absence of analysis of the relevance of content according to age, systematic measurement of adverse effects, tools to limit screen time, and the near absence of documented ethical standards for working with young people in vulnerable situations. (p.16–19)

6. FALC Version 

6.1 FALC Summary

What is this text about?

• The text talks about robots that chat with young people.

• They are called “chatbots” or “conversational agents”.

• They talk about mental health, for example sadness, stress, anxiety.

Who are the young people?

• They are individuals aged between 15 and 24.

• Some are doing well, others have mental health issues.

What the researchers did

• They studied 39 research papers on these chatbots.

• Most of the research is recent.

• Young people use chatbots for a short time, often less than a month.

What young people like

• Talking to the chatbot whenever they want, day and night.

• Having a kind and understanding tone.

• Being able to talk about difficult topics anonymously.

What young people do not like

• The responses often repeat.

• The chatbot does not always understand sentences well.

• The responses are not always suitable for their situation.

What the text says for adults

• Chatbots can help a little, but do not replace professionals.

• We must protect young people: their data, their emotions, their screen time.

• We must build these tools with young people and with caregivers.

6.2 Key points FALC

1. The studies are short

• The tests last a few days or a few weeks.

• We do not know what happens in the long term.

2. Chatbots are still in testing

• Many tools are prototypes.

• They mainly use pre‑written texts.

3. What are chatbots for?

• To talk about sadness, stress, anxiety.

• Sometimes about body image, alcohol or phone.

4. What works well

• It is easily accessible on phone or computer.

• Young people sometimes feel less alone.

5. What poses a problem

• The content is sometimes poor and repetitive.

• There is little discussion about security and data protection.

7. Cross-sectional analysis – Values of health practices

• Literacy : the document notes that many young people find the texts too long or too technical and recommends shorter, multimedia content that is adapted to different levels of understanding. (p.12, 16–18)

• Empowerment : several chatbots aim at skill acquisition (CBT, self‑compassion, problem-solving), but the co-construction of content with young people remains limited and often secondary. (p.8–10, 13, 18–19)

• Participation : some studies describe a co-design with young people, parents and professionals, but this concerns a minority; the participation of young people is not yet a norm. (p.5, 7–9, 18–19, 31)

• Community health : some interventions target specific groups (LGBTQ+, indigenous communities, refugees) and mobilise community actors, but the structured community health logic remains underdeveloped. (p.7–9, 17)

• Ethics : the authors highlight the almost general absence of structured reflection on cultural, social or technological biases, and on the management of crisis or vulnerability situations. (p.6–7, 16–19)

• Human rights : the text emphasises the importance of confidentiality, security and equitable access, while showing that most studies do not precisely document these aspects. (p.6–7, 16–19)

• Intersectorality : several works involve schools, universities, health services and associations, but partnerships often remain sporadic, without a formalised intersectoral framework. (p.7–9, 27–33)

• Partnership : the review clearly calls for structured collaborations between young people, clinicians, researchers, developers and decision-makers, but notes that these partnerships are still rare. (p.18–20)

• Combating discrimination : some initiatives explicitly target LGBTQ+ youth or cultural minorities, but the issue of discrimination, non-judgment and diversity is insufficiently addressed and not systematised in the studies. (p.7–9, 17, 50)

8. Assessment of the reliability of the resource

Scientific relevance

• Strengths :

  • Rigorous scoping review method, following PRISMA, with searches across four major databases, clear inclusion/exclusion criteria, and structured thematic analysis. (p.3–5)

  • Detailed mapping of study characteristics, design choices, and technical and evaluative dimensions, supported by tables and boxes. (p.6–15)

  • Argumentative discussion of the methodological limitations of the studies (small sample sizes, short durations, few safety evaluations). (p.17–20)

• Limitations :

  • Dependence on information reported in primary articles, which may underestimate certain aspects of design or safety that are not described. (p.20)

  • Absence of quantitative analysis such as meta-analysis (normal for a scoping review, but limits quantifying clinical effectiveness). (p.1–2, 14–18)

Operational relevance

• Strengths for the field :

  • Provides a comprehensive framework for evaluating or designing a chatbot: target audience, objectives, architecture, AI used, content, usage modalities, strengths/weaknesses, ethical aspects. (p.4–5, 10–15, 18–19)

  • Offers directly usable design guidelines as a basis for specifications or usage charters. (Textbox 3, p.19)

  • Highlights crucial blind spots (safety, age, ethics, vulnerability) that can serve as a checklist for vigilance for organisations. (p.16–19)

• Limitations for the field :

  • Few “ready-to-use” examples of integration into concrete public health or routine prevention measures. (p.17–20)

  • The majority of the tools described are not mature enough for wide deployment without local adaptation and evaluation. (p.10–13, 17–20)