Co-Design Is Not Evidence of Efficacy for Mental Health AI

The conversational role does not arise in the computational model.

Grové's contribution addresses a system that was developed together with young people, technology partners, and professional stakeholders. Part of it works with artificial intelligence and rule-based methods of natural language processing. The intention is to convey evidence-based resources, information about mental health, support for well-being, and adaptive coping strategies. This is a clearly more limited description than the notion of a freely acting digital expert.

The central subject is not the performance capability of a model, but the social form of the offering. According to the publication, interviews and surveys fed into the system's personality and character; in addition, conversational design and content development are addressed. Young people thus appear not merely as a later user group, but as participants in defining how the system speaks and presents itself. From an editorial perspective, we consider this shift essential: In sensitive conversations, the role is not a decorative overlay on a technical function. It influences what expectations the product generates in the first place.

Grové documents development, not treatment success.

The research achievement of the lead publication lies in the participatory development process. It asks, in essence, how a digital mental health and wellbeing offering can be designed with and for young people, and what lessons or reservations arise from this. The text also discusses a possible use in secondary schools or health facilities – explicitly in connection with the well-being or care teams there. This speaks against the reading of an isolated self-help machine.

However, the provided information does not reveal sample sizes, recruitment, precise evaluation methods, or controlled comparisons. Nor are clinical endpoints, usage periods, or safety results reported. Therefore, it cannot be inferred from this source that the offering reduces anxiety, depressive symptoms, or school-related stress. Acceptance or long-term use are also not substantiated by the available information.

This limitation does not diminish the contribution as long as it is read as a development report. It would only become problematic if a successful participatory process were presented as evidence of efficacy. Co-design can show which language, character, or conversation structure young people prefer and which content appears relevant to them. On its own, it cannot determine whether the finished system reliably helps, harms, or has no consequences in stressful situations.

Fit is a distinct but limited type of evidence

Participation can address a real development problem: adult professionals and product teams do not automatically know how young people perceive a digital counterpart. Content that is technically correct can fail due to an unsuitable character, a lecturing tone, or an implausible conversation style. Grové's approach is therefore more than a friendly design gesture. It generates knowledge about the intended usage situation that does not come from technical tests.

Nevertheless, perceived fit must not be conflated with health effects. A likable character can facilitate use; whether it actually does so would need to be investigated separately. Even frequent or prolonged use would not be proof of a positive health outcome. Our editorial position is therefore: co-design belongs in the evidence chain, but at its beginning. It establishes the relevance and acceptability of a concept, not its clinical claim.

The system's personality is not a side issue

The fact that interviews and surveys influenced the personality and character design points to an underestimated area of product development. A digital system does not speak neutrally. Word choice, response length, handling of uncertainty, and the presentation as a character suggest whether it informs, accompanies, instructs, or claims authority. The more human and approachable this presentation appears, the more important a precise boundary becomes between conveyed closeness and actual responsibility.

The source reports examples of conversation and content development, but the provided summary does not allow a detailed examination of individual dialogues. It thus remains open how the system handles ambiguity, acute distress, or misunderstood inputs. This knowledge boundary should not be filled with speculation. From the described goal, it only follows that resources, information, and coping strategies are to be communicated – not that every response is technically correct, situationally appropriate, or safe.

NLP research often investigates a different problem

The systematic review by Le Glaz and colleagues broadens the perspective on machine learning and natural language processing in mental health. After a search in four medical databases, 58 of 327 identified articles were included in a qualitative analysis. The studies examined were methodologically and thematically heterogeneous. Their goals included extracting symptoms, classifying disease severity, comparing treatment effectiveness, and deriving psychopathological indicators.

This research landscape does not fully align with a conversational offering for adolescents. Much of this research concerns medical records or social media and thus the analysis of existing language, not the conduct of a supportive conversation. The review also notes that powerful classifiers were preferred over transparent models. This creates a conflict for a conversational system: technical predictive performance can be relevant, but in a sensitive application context, responses must also be limited in a comprehensible way, and the organization must take responsibility for them.

Le Glaz and colleagues also warn against the inaccuracy of social media users as a cohort and point to language-specific features. This is particularly significant for youth services. A system that works in one language or on the basis of certain data types cannot simply be transferred to other linguistic and cultural contexts. Grové's local participation logic and the methodological breadth of the review thus point to the same task: context must be investigated, not merely asserted.

Generative AI increases the burden of justification

Torous and Blease turn their attention to generative AI in 2024. As far as the provided text indicates, they do not report their own randomized intervention study, but rather assess possible applications and current problems. They distinguish administrative support, documentation, training, and symptom monitoring from the considerably more demanding questions of prevention, diagnosis, and treatment. Research on non-clinical samples has partly focused on perceived empathy rather than clinical outcomes.

This distinction sharpens the assessment of co-design. A system perceived as empathetic can be a communicative success without having a proven effect. Torous and Blease also point to a lack of evidence for guiding mental health treatment, insufficient data protection, potential biases, and the danger of errors that are hard to detect. For clinical claims, they demand high-quality randomized controlled trials, including appropriate digital comparators. This is a different level of evidence from the development of personality and dialogues.

Embedding determines the real significance

Grové's consideration of deploying the offering together with school or health teams is therefore more than a question of distribution channel. The environment determines who explains the system, who receives feedback, and how its limited role becomes apparent. The publication does not, however, demonstrate that such an embedding has already been successfully implemented or evaluated. It articulates a possible context, not an implementation success.

Torous and Blease argue similarly against permanently isolated programs. In addition to evidence and data protection, they emphasize participation, clinical integration, and interoperability. This is not only about technical connections. Professionals must also be able to actually use a system, and workflows, regulation, and qualification influence its introduction. A well-designed youth offering can therefore fail due to institutional reality, even if its character is well received. Conversely, integration does not automatically make an untested system safe.

After co-design, the actual examination begins.

From the three publications, no complete evaluation model emerges, but rather a clear sequence of open questions. First, an offering must be designed to be understandable and relevant for its target group. Then, technical reliability, data protection, biases, and linguistic-cultural transferability must be examined. If health effects are promised, studies are ultimately needed that go beyond agreement, perceived empathy, or model performance and capture actual outcomes as well as potential harms.

Our editorial thesis remains deliberately strict: youth participation is a prerequisite for responsible design, but not a seal of quality for the finished product. Grové's contribution is valuable precisely when one does not overload it. It shows how future users can participate in the voice and form of a system. Whether this system is useful in the school or care context, remains safe, and can be sustainably integrated is not answered by this. A serious next publication would have to not only tell how the conversation was designed, but also examine what happens after this conversation.