Debate: Conversational AI and young people's mental health – friend or foe? – conversational AI and young people's mental health: the case for clinically informed design

Abstract

Conversational AI is increasingly being used by young people for emotional support, advice and conversations about personal concerns. Emerging evidence suggests that these systems may offer mental health benefits, particularly for depressive symptoms, while also raising concerns about dependency, relational dynamics, safety and therapeutic precision. This debate article argues that the central question is not whether conversational AI should be viewed as beneficial or harmful, or as therapeutic or nontherapeutic. Rather, its impact on young people's mental health may depend substantially on how these systems are designed. As the boundaries between clinically oriented and general‐purpose conversational agents become increasingly blurred, design decisions concerning relational style, responses to distress, risk management, transparency and human oversight take on growing importance. These decisions are clinical and developmental as much as technical. The article argues that mental health expertise has an important role to play in shaping, evaluating and governing conversational AI systems, particularly as they become more embedded in the everyday lives of young people.

Any claim about AI‐based conversational agents risks proving premature or quickly outdated. This is a powerful technology whose widespread adoption is still recent, whose evolution is unusually rapid and whose evidence base remains in its early stages. With so few certainties available, it may be more useful to consider emerging evidence of benefit alongside early signals of risk, paying particular attention to how these systems are designed, deployed and used by young people.

The potential benefits seem clearest where these systems are designed for therapeutic purposes, or at least for emotional support. Early evidence suggests that conversational agents may have therapeutic value, particularly for reducing depressive symptoms in young people. Recent reviews and meta‐analyses report promising short‐term effects, although findings are less consistent for anxiety, wellbeing and longer‐term outcomes (Feng et al., 2025; Feng, Tian, Ho, Yorke, & Hui, 2025).

At the same time, this format seems to address one of the most persistent difficulties in digital mental health: engagement. The level of personalisation that conversational agents can achieve, together with their capacity to sustain friendly and empathic conversations about sensitive topics, may make them not only relevant and accessible, but also attractive and even enjoyable to use (Hou et al., 2025). For years, the field has asked how to encourage young people to actually use digital interventions and how to reduce high rates of attrition. Rule‐based chatbots now feel almost distant, even though only a few years have passed since they were the norm. We now find ourselves asking whether this new generation of tools may, in fact, produce too much engagement.

This is where the promise becomes more complicated. The same features that make conversational agents engaging may also create new risks: dependency, the substitution of human relationships in more vulnerable adolescents, reduced therapeutic precision and the challenge of designing systems capable of responding safely to risk (Kim, Xie, & Yang, 2025; Wu et al., 2026). One might be tempted to manage these concerns by drawing a clear distinction between agents designed for therapeutic purposes and those that are not: assuming that the former have the necessary guardrails to address these challenges, while the latter do not because they serve a different purpose. Yet this distinction is becoming increasingly difficult to sustain. Young people are turning to general‐purpose AI systems for advice, emotional support and conversations about personal concerns, regardless of whether those systems were originally designed for mental health purposes (McBain et al., 2025). Conversely, a growing range of commercially available conversational agents are presented as scientifically informed, or as having a therapeutic layer, despite often relying on limited independent evidence regarding their safety and effectiveness (Common Sense Media, 2025).

This blurring of boundaries suggests that the central question might not be whether conversational AI should be classified as beneficial or harmful, or therapeutic or nontherapeutic, but whether it is designed with enough clinical, developmental and relational awareness for the ways young people may actually use it (Mansoor, Hamide, & Tran, 2025). This shifts the debate from the technology itself to the design decisions that shape the interaction. These decisions include how the system presents itself to users, how it responds to distress, what role it is intended to play in a young person's life and how it handles situations that require human intervention.

Good design might not look the same in every case, because conversational agents may occupy very different roles in young people's lives. Agents explicitly designed for therapeutic or emotional support may need a clear rationale, conversations guided by an explicit clinical framework and appropriate human oversight or human‐in‐the‐loop mechanisms. General‐purpose systems may not be designed to deliver treatment, but they can still benefit from choices that promote wellbeing, such as maintaining appropriate boundaries, reducing anthropomorphic framing and avoiding the reinforcement of harmful beliefs. Across this spectrum, the emerging literature repeatedly highlights the importance of transparency about the system's capabilities and limitations, mechanisms for recognising and responding to risk, signposting to human support when needed and careful consideration of how the agent positions itself in relation to the user (Meadi et al., 2025).

Design choices in this space inevitably carry clinical and developmental implications. When should an agent validate a young person, and when should it gently challenge them? How much relational warmth is appropriate? When should a disclosure raise concern? How should the system move a user towards human support? These questions require expertise in mental health, risk and adolescent development. Recent initiatives from general‐purpose AI companies illustrate this shift: Anthropic, for example, has described work on suicide and self‐harm classifiers, crisis signposting, collaboration with suicide prevention experts and reducing sycophancy in conversations where users may need challenge rather than simple agreement (Anthropic, 2025). This does not mean that every conversational AI system should be treated as a clinical intervention. But when a system is likely to be used by young people to ask for advice, share emotions, seek reassurance or make sense of personal difficulties, mental health expertise becomes increasingly important in shaping how the system is designed, tested and governed.

Of course, clinically informed design is not a guarantee of benefit, and a design‐centred argument should not be mistaken for design optimism. Generative systems remain partly unpredictable, commercial pressures may conflict with wellbeing and young people often use technologies in ways that designers did not anticipate. But the future impact of conversational AI on youth mental health may depend less on advances in the technology itself than on whether clinical, developmental and engineering expertise can work together to shape how that technology enters young people's lives.

For mental health professionals, this may represent an important shift. New roles are emerging that are not a simple extension of traditional clinical expertise. Contributing meaningfully to conversational AI will require understanding how product decisions are made, how clinical judgement can be translated into design requirements, how to work with engineers and product teams, and how to evaluate systems that may change over time. Yet collaboration with industry goes beyond these practical challenges. It also requires attention to commercial incentives, data access arrangements, transparency and conflicts of interest that may shape how these systems are built and studied. This kind of industry literacy is likely to become an important part of responsible clinical engagement with technology companies (Dubicka et al., 2025). If conversational AI is to play a growing role in young people's lives, mental health professionals may need to become more active participants in the difficult work of making these systems clinically and developmentally safer.

Funding information

No funding was received for the preparation of this article.

Conflict of interest statement

The author has worked with startups and nonprofit organisations involved in the design, development and evaluation of conversational AI systems and digital mental health interventions for young people, and this experience has informed some of the perspectives discussed in this article. The author holds no equity or financial interests in any organisation that may benefit from the views expressed.

Ethics statement

No ethics approval was required for this debate article.

Data availability statement

Data sharing is not applicable to this article, as no datasets were generated or analysed for this debate article.