Chatbots reduce distress in young people by a small but measurable amount

29 interventions form a heterogeneous field

The authors systematically searched eleven databases and search engines. Studies from more than ten years were included in which chatbots were used to improve the mental health of young people.

13 of the 29 studies were randomized controlled trials and were included in quantitative analyses. The remaining studies contributed to the narrative assessment of feasibility and acceptability.

The sheer number of influencing factors shows that chatbot does not denote a uniform intervention. Rule-based systems, structured exercises, and generative responses can function very differently under the same interface.

A meta-analysis usefully summarizes results but cannot fully resolve the differences. The average effect does not automatically apply to every product and every target group.

Psychological distress decreased with a small effect

For distress, the meta-analysis found a Hedges g of -0.28 with a 95% confidence interval from -0.46 to -0.10. The effect was statistically significant and small in magnitude.

A small effect can be practically relevant for an easily accessible and scalable offering, especially when it reaches many people. However, it must not be promised as a strong individual effect.

The comparison context is also crucial. A chatbot compared against a waiting list, informational material, or active human support answers a different question in each case.

For a specific individual, the benefit may be greater, smaller, or nonexistent. Averages describe populations, not predictions for individual cases.

No significant effect on well-being

For psychological well-being, Hedges g was 0.13, with a confidence interval ranging from -0.16 to 0.41. The effect was therefore not statistically significant.

Reducing distress and increasing well-being are different goals. A system may somewhat alleviate symptoms or acute distress without measurably changing life satisfaction, connectedness, or positive development.

Product communication should not summarize these goals. "Does good" can mean very different things and can only be substantiated with clear outcome measures.

The null result is not proof that chatbots never influence well-being. It shows that the aggregated evidence in this analysis could not demonstrate a reliable average effect.

Design and delivery influenced the effect

Subgroup analyses showed differences by sample type, delivery platform, interaction mode, and response generation method. Effect thus depends on more than just the chatbot label.

Instant messenger platforms are recommended as a potentially suitable delivery format. Familiar channels can facilitate access and use, but they bring their own data protection and platform dependencies.

Multiple communication modes can improve interaction quality. Whether audio, image, or video are actually necessary should be weighed against data minimization and the specific target group.

Generative and pre-structured responses also have different strengths. Free language increases flexibility, while curated modules facilitate control and reproducibility.

Woebot provides an early example

Fitzpatrick, Darcy, and Vierhile compared two weeks of Woebot with an informational e-book in 70 young adults. The chatbot offered CBT-based self-help in up to 20 sessions.

The Woebot group showed a greater reduction in PHQ-9 scores in the intention-to-treat analysis. In the complete cases, anxiety scores decreased in both groups.

Participants used the chatbot on average 12.14 times. Comments suggested that process factors shaped acceptance more strongly than therapeutic content alone.

The study shows a positive short-term signal, but it is small, unblinded, and short. The meta-analysis helps to place such individual findings into a broader, still heterogeneous field.

A youth pilot shows the limits of small studies

Nicol and colleagues randomized 18 adolescents with moderate depressive symptoms to a CBT app with a conversational agent or to a waitlist. 17 individuals were included in the analysis.

The mean PHQ-9 decreased by 3.3 points in the app group and by two points in the control group after four weeks. Acceptability, feasibility, and usability were rated positively.

The authors emphasize that the pilot could not provide evidence of efficacy. The predominantly female, white, and privately insured sample further limits generalizability.

Such pilot data are important for planning and implementation. They should not be communicated with the same certainty as a larger, adequately powered randomized study.

Feasibility and acceptability remain outcomes in their own right

The review assesses chatbot interventions overall as feasible and acceptable. That is a necessary precondition for efficacy, but it is not identical to efficacy.

An offering may be accepted in a study because support, onboarding, and clear tasks are in place. In open everyday use, registration, costs, technical errors, or mismatched conversations can change usage.

Acceptability should also not be measured only as satisfaction. Understanding of the AI’s role, data processing, the possibility of correction, and the freedom to discontinue are all part of it.

For young people, it is especially important which groups are not reached. Average good acceptability can conceal linguistic, social, or technical exclusions.

Moreover, usage should not automatically count as agreement with the entire concept. Adolescents may find a single exercise helpful and at the same time reject the chat style. Component-level feedback helps avoid blending efficacy, content, and interface into a vague overall impression.

Dropout rates must be supplemented with reasons. An early exit can mean disinterest, technical problems, a lack of fit, or simply a short-term purpose already achieved. Without this information, engagement is hard to interpret.

The evidence supports supplementation with clear scrutiny

The meta-analysis supports a small positive effect on psychological distress in young people. For well-being, it found no significant effect, and the impact depended on several design factors.

The authors view chatbots as a possible supplement to existing services provided by multidisciplinary professionals. They recommend improvements in language processing, accuracy, privacy, and security.

For new generative conversational partners, it is not enough to refer to older CBT chatbot studies. The model, prompt, conversation trajectory, and open-ended responses change the intervention and require their own testing.

Likewise, costs and accessibility should be reported together. An inexpensive service can supplement care, but technical limitations, wait times, or a suddenly unavailable chat affect the real-world benefit for young people.

Chatbots can reduce distress by a small but measurable amount. This claim only becomes credible with its full context: average effect, limited outcome measure, heterogeneous systems, and the need for subsequent independent research.