Using Generative AI to Co-Design Digital Mental Health Interventions With Adolescents in Rural South Africa: Qualitative Thematic Analysis of Participatory Workshops

Background: Digital mental health interventions (DMHIs) offer a scalable approach to address adolescent depression and anxiety. User-centered coproduction can optimize acceptability and engagement, but it is often resource-intensive. Advances in generative artificial intelligence (GenAI) create new opportunities for involving adolescents in co-design, yet research on its feasibility and acceptability, particularly in low-resource settings, remains underexplored. Objective: This study aimed to explore adolescents’ experiences and perspectives of using GenAI to co-design stories, images, and music for the Kuamsha app (Sea Monster), a gamified DMHI that teaches behavioral activation through interactive narratives and peer support. Methods: Overall, 2 participatory workshops and focus group discussions were conducted with 23 adolescents (aged 15‐19 years) in rural Mpumalanga, South Africa. Participants were guided to use 3 GenAI tools—ChatGPT (OpenAI), text-to-story; MidJourney (MidJourney Inc), text-to-image; and Soundful (Soundful Inc), music generation—to create digital content. Data were audio-recorded, translated, transcribed, and triangulated with the facilitator’s observation notes. Thematic analysis was used to explore key themes. Results: Almost all participants (22/23, 96%) had no prior exposure to GenAI. The majority (20/23, 87%) described the creative process as enjoyable and engaging, with most (21/23, 91%) reporting that creating music improved their mood. Adolescents expressed autonomy and ownership of the process, with more than half (14/23, 61%) personalizing outputs to reflect their identities and aspirations. All participants (23/23, 100%) preferred artificial intelligence (AI)–generated images over the cartoon-like illustrations of the Kuamsha app, and most (19/23, 83%) preferred AI-generated music. Story preferences were more mixed, with about a quarter of participants (6/23, 26%) recalling that Kuamsha’s narratives contained embedded lessons that were not integrated into the ChatGPT outputs. Most adolescents (18/23, 78%) required support with prompt construction, and more than half (13/23, 57%) noted cultural biases in AI outputs, particularly in images. Most participants (17/23, 74%) expressed interest in using AI for schoolwork and creative projects, while a minority (6/23, 26%) preferred to limit use to personal applications. Concerns about fairness and the displacement of human creativity were also raised. Conclusions: GenAI shows promise for enhancing adolescent engagement in the coproduction of DMHIs and enabling culturally relevant and personalized content. However, reliance on human support and persistent algorithmic biases remain limitations. Further research should explore the integration of therapeutic principles into AI-generated media and strategies to mitigate bias.