Lived Experience in Dialogue: Co-designing Personalization in Large Language Models to Support Youth Mental Well-being

Abstract: Youth increasingly turn to LLMs for mental well-being support, yet current personalization in LLMs can overlook the heterogeneous lived experiences shaping their needs. We conducted a participatory study with youth, parents, and youth care workers (N=38), using co-created youth personas as scaffolds, to elicit community perspectives on how LLMs can facilitate more meaningful personalization to support youth mental well-being. Analysis identified three themes: person-centered contextualization responsive to momentary needs, explicit boundaries around scope and offline referral, and dialogic scaffolding for reflection and autonomy. We mapped these themes to persuasive design features for task suggestions, social facilitation, and system trustworthiness, and created corresponding dialogue extracts to guide LLM fine-tuning. Our findings demonstrate how lived experience can be operationalized to inform design features in LLMs, which can enhance the alignment of LLM-based interventions with the realities of youth and their communities, contributing to more effectively personalized digital well-being tools.

• The paper demonstrates a co-design methodology that integrates lived experience to personalize LLMs and improve youth mental health support.
• It employs a rigorous four-stage participatory process, refining personas with youth and care workers to ensure dialogic scaffolding and explicit safeguards.
• Findings challenge rule-based DMHI approaches by advocating dynamic, context-sensitive methods that promote user autonomy and psychological safety.

Co-designing LLM Personalization for Youth Mental Well-being: Embedding Lived Experience

Introduction

This paper, "Lived Experience in Dialogue: Co-designing Personalization in LLMs to Support Youth Mental Well-being" (2511.05769), systematically investigates how LLM-based digital mental health interventions (DMHIs) can operationalize lived experience to improve youth well-being support. Leveraging co-design with youth, parents, and youth care workers, the study advances a person-centered framework for LLM personalization, foregrounding dialogic and contextual adaptation, explicit system boundaries, and scaffolding for reflection and autonomy. The research addresses limitations of current rule-based DMHIs, critiques prevailing models of “personalization,” and demonstrates the methodological and practical value of participatory persona-driven design for LLM-based well-being tools.

Limitations of Rule-based and Generic LLM Personalization

Prevailing DMHIs typically utilize categorical, rule-based personalization (age, diagnosis, or demography-dependent tailoring), which fails to capture the complexity and dynamism characterizing lived youth experience. Such methods risk reductive generalization, poor developmental fit, and lack of psychological safety. The study critiques commercial LLM-chatbots and generic platforms—frequently adopted by youth without oversight—for their lack of evidence-based design, inappropriate scope (e.g., handling crisis), and misalignment with the nuanced values and contexts of adolescents and their communities. The authors note increased risks (hallucinations, algorithmic bias, overreliance, and failure to provide appropriate referral pathways), especially in absence of co-designed safeguards.

Participatory Persona-based Co-design Methodology

The authors designed a four-stage participatory process:

1. Data-driven Persona Scoping: Leveraging survey clustering and thematic analysis of public youth helpline forums, the study constructed 10 quantitatively-derived and 3 qualitatively-derived personas representing diverse well-being strategies and challenges.
2. Lived Experience-based Persona Refinement: Three iterative co-creation workshops with marginalized youth refined these personas, enriching static profiles with nuanced daily routines, socio-relational contexts, emotional trajectories, and digital habits, ultimately producing 7 participatory personas.
3. Stakeholder Validation and Derivation of Personalization Priorities: 14 semi-structured interviews with youth, parents, and youth care workers used these participatory personas to elicit core requirements, boundaries, and mechanisms for personalization.
4. Translation to LLM Design and Dialogue Extracts: Inductive thematic analysis (supplemented by Persuasive System Design—PSD—feature mapping) produced actionable dialogue scaffolds for LLM fine-tuning.

This methodology directly confronts the epistemic challenges of integrating lived experience into computational system design, embedding the voices of youth as “experts by experience” rather than downstream evaluators or respondents.

Core Themes: Redefining Personalization

Three cross-cutting themes emerged as foundational for effective LLM personalization in youth DMHI:

1. Person-centered Contextualization

Stakeholders emphasized ongoing adaptive understanding over static demographic or diagnostic tailoring. Effective personalization requires continuous probing for context—educational background, dynamic routines, relational dynamics (beyond roles), digital and offline habits, and the unique, fluctuating manifestation of challenges. Nuanced causes of distress (e.g., distinguishing financial strain due to gambling vs. family pressure) demand interpretive, dialogic interaction.

2. Explicit Boundaries and Safeguards

All groups stressed unequivocal boundaries: LLMs must reinforce their non-human status, avoid clinical/diagnostic advice, and operate within clearly defined scopes. Critical safeguards include transparent data handling, explicit consent and privacy mechanisms, visible and immediate referral pathways to offline support, and the strategic leveraging of peer narratives (without veering into overidentification). Neutral tone and mitigation of simulated empathy were repeatedly highlighted, counteracting the risk of overreliance or parasocial attachment.

3. Dialogic Scaffolding for Reflection and Autonomy

Rather than deliver prescriptive advice, LLMs must implement dialogic scaffolds—stepwise interpretive questioning, intent clarification, facilitation of narrative exploration, and autonomy-preserving suggestions. Systematic support for youth self-reflection is prioritized over “solution provision.”

Operationalization: From Themes to System Features and Dialogue

The translation of these principles into system design utilized PSD feature mapping augmented by direct-from-stakeholder dialogic strategies. For example:

• Primary Task Support (PTS): Tailoring by inquiring about user context; reduction and tunneling for complex goal breakdown.
• Social Support (SS): Peer narrative retrieval; focus on relational experience over taxonomy of social roles.
• System Credibility Support (SCS): Reinforced through transparency, explicit behavioral boundaries, and real-world references.
• Dialogic Strategies: Open-ended “why” questions, intent clarification at each exchange turn, persistent autonomy prompts, and minimization of affective mimicry.

Fine-tuning data constructed from these dialogue extracts and strategic interventions is positioned to move LLM-based DMHIs beyond crowdsourced, generic, or professional-only annotation paradigms.

Theoretical Implications

The study advances HCI theory by reframing personalization in digital health as a dialogic, context-sensitive process subject to ongoing co-adaptation, rather than a rules-based tailoring exercise. The results challenge extant persuasive system design taxonomies, arguing for expanded recognition of dialogic scaffolding and lived context as central to effective, safe personalization. Participatory personas authored by marginalized youth function as epistemic artifacts—bridging the divide between lived context and technical implementation.

Practical Implications for Researchers, Developers, and Policymakers

1. DMHI Researchers and Builders: The creation of fine-tuning pipelines incorporating community-derived dialogue extracts can incrementally align LLMs to youth well-being priorities. Dialogic scaffolds and explicit boundaries are to be treated as engineering requirements, not “add-ons.”
2. Evaluation Infrastructure: Results underscore the importance of participatory continuous evaluation loops (design, model training, real-world testing). Reliance on general crowd-annotators is insufficient; lived experience experts are required for annotation and evaluation.
3. Policy and Governance: The study demonstrates the strategic value of community-based co-design, both for technical safety and for aligning with the developmental and sociocultural realities of marginalized youth. Investment in participatory infrastructure is positioned as a foundation for responsible innovation and digital health equity.

Limitations and Future Directions

The primary limitation pertains to the validation stage: it remains to empirically verify whether the derived personalization strategies and dialogue scaffolds are internalized effectively by LLMs post-fine-tuning. Further research is required to examine generalizability outside the Dutch context, address potential creation-phase biases in participatory personas, develop dynamic persona scenarios for continuous training, and extend application to clinical/specialist domains.

Conclusion

This paper demonstrates that fine-tuned, person-centered personalization for LLM-based youth DMHIs is feasible and that lived experience, systematically operationalized through participatory persona co-design and stakeholder dialogue, fundamentally augments both safety and efficacy. The work repositions personalization as a process of dialogic inquiry, contextual adaptation, and autonomy support—directly challenging current practices of static tailoring. Future lines of investigation should refine participatory fine-tuning pipelines and undertake rigorous evaluation of the impact of lived experience-derived alignment on real-world youth mental well-being outcomes.