TLDR: A study co-designed personalization strategies for Large Language Models (LLMs) to support youth mental well-being, involving youth, parents, and youth care workers. It identified three key themes: contextual understanding, safety boundaries, and dialogic scaffolding for reflection and autonomy. These insights were translated into concrete LLM dialogue extracts and design safeguards, advocating for a person-centered approach to make AI interventions more relevant, safe, and empowering for young people.
Large Language Models (LLMs) are increasingly becoming a go-to resource for young people seeking support for their mental well-being. However, the way these AI systems currently personalize interactions often misses the mark, failing to account for the diverse and complex real-life experiences that shape a young person’s needs. A recent study titled “Lived Experience in Dialogue: Co-designing Personalization in Large Language Models to Support Youth Mental Well-being” addresses this crucial gap by bringing together youth, parents, and youth care workers to collaboratively design more meaningful and safe personalization strategies for LLMs.
The research, conducted by a team including Kathleen W. Guan, Sarthak Giri, Mohammed Amara, and others, highlights an urgent need for better interventions to support youth mental well-being. With a significant portion of mental health disorders emerging before age 25, and prevalence rates on the rise, digital platforms offer accessible and anonymous support. While custom chatbots are already used for emotional check-ins and psychoeducation, many struggle to be truly relevant and responsive to the dynamic needs of young individuals.
Traditional personalization in digital mental health interventions (DMHIs) often relies on rule-based approaches, using fixed categories like demographics or symptom scores. While efficient, this method frequently overlooks the nuanced and evolving identities and circumstances of youth. LLMs, with their ability to generate adaptive responses based on conversation, offer a promising alternative for more fluid and personalized interactions. Yet, their novelty also brings risks such as overreliance, algorithmic bias, and hallucinations, especially concerning for vulnerable youth.
To address these challenges, the researchers adopted a participatory design approach, emphasizing co-creation with youth and their support communities. This ensures that LLM responses reflect the values and lived realities of young people, rather than just assumptions from training data or professional guidelines. The study focused on preventative well-being, aiming to inform the design of LLMs that enhance support while mitigating potential harms.
A Four-Stage Journey to Person-Centered Design
- Scoping, data-driven personas: Initial personas were created from existing survey data on coping strategies and youth forum posts, serving as a starting point for discussion.
- Lived experience-based co-creation: Youth participated in workshops to critique and refine these preliminary personas, creating “participatory personas” that truly reflected their own experiences, daily routines, stressors, and digital habits.
- Community-based interviews: These youth-created personas were then used as discussion tools in interviews with other youth, parents, and youth care workers. Interviewees provided insights on how LLMs could personalize support safely and effectively.
- Translation to design: The insights from these interviews were analyzed thematically and mapped to persuasive system design (PSD) features, resulting in concrete dialogue extracts and design safeguards for LLM fine-tuning.
Key Insights for LLM Personalization
The analysis revealed three overarching themes for designing personalized DMHIs for youth:
1. Contextual Understanding as the Core of Personalization: Stakeholders emphasized that personalization must go beyond simple demographic labels. It requires LLMs to continuously probe for the underlying meaning and contextual causes of a young person’s challenges. This includes understanding their educational background, the nuanced reasons behind common problems (e.g., why they avoid school), the dynamics of their relationships (not just who is present, but how those relationships feel), and their everyday routines and digital practices. Crucially, LLMs need to adapt dynamically to the fast-changing contexts of youth, ensuring advice remains relevant as situations evolve.
2. Safety Boundaries in Personalization: Participants stressed the importance of clear limits to prevent harm. Key concerns included privacy and disclosure, with youth needing assurance that their shared information would remain confidential. Transparency and role clarity were vital, meaning LLMs should explicitly state their limitations (e.g., “I am not a human”) and explain how data is handled. Strict content and behavioral boundaries were called for, avoiding overconfident responses on sensitive topics and mitigating risks of overreliance. When appropriate, LLMs should frequently refer users to offline human resources and support. Interestingly, the use of peer narratives was seen as valuable for grounding advice in relatable experiences, and a neutral tone was preferred over overly compassionate or empathetic mimicry, which could feel inauthentic.
3. Dialogic Scaffolding for Reflection and Autonomy: Personalization was viewed as a process of structured dialogue that empowers youth to think through their situations, rather than simply receiving ready-made answers. This involves intent clarification (asking what the user wants to discuss), interpretive questioning (exploring personal meaning), and narrative and goal exploration (linking short-term challenges to broader aspirations). The goal is to support autonomy, guiding reflection and pattern recognition while leaving final choices to the young person.
Translating Insights into Actionable Design
These themes were translated into practical LLM design strategies. For instance, to achieve “Contextual Understanding,” LLMs could start conversations by asking, “What are the most important things I should know about you in order to help you?” or offer responses in varying literacy levels. For “Safety Boundaries,” dialogue extracts were created to clarify data handling, state system limitations, and provide referrals to helplines like Kindertelefoon. To foster “Dialogic Scaffolding,” strategies included frequently asking “why” to encourage deeper reflection and ending each LLM turn with a suggestion followed by an invitation for user feedback or alternative exploration.
Also Read:
- Simulating Learners: How AI is Reshaping Educational Research and Practice
- Unlocking Cognitive Distortions: An Interpretable AI Approach
Implications for the Future of AI in Mental Health
This study offers significant implications for digital health research and development teams. It advocates for framing personalization as a continuous, dialogic process grounded in lived experience, moving beyond static tailoring. The identified dialogic scaffolds provide concrete interactional features for future studies to evaluate. For LLM developers, the findings emphasize integrating these personalization priorities into training data and system pipelines, moving away from crowdsourced annotators towards input from lived experience experts, including youth themselves.
For public health policymakers, the research underscores the value of investing in community-based co-creation. Involving youth, parents, and youth care workers in the design process helps define the boundaries of digital health personalization, clarifies appropriate referral mechanisms, and ensures digital tools complement, rather than replace, human care. This collaborative approach can strengthen the capacity to monitor, evaluate, and govern the integration of LLMs in community and care settings, ultimately leading to more equitable AI systems for youth mental health.
While the study acknowledges limitations, such as the need to empirically validate the effectiveness of these fine-tuning guidelines and the localized sample, it lays crucial groundwork. It demonstrates a structured approach for embedding lived experience into LLM personalization design, fostering adaptive support that promotes reflection and autonomy in youth mental well-being.
