JMIR Publications

ABSTRACT

Background:

Crisis text lines provide vital mental health support to youth, yet scaling training and quality assurance for a large, diverse volunteer workforce to ensure confident, standards-aligned care remains a significant challenge. Generative AI offers potential for realistic practice conversations, but current approaches lack empirical grounding in authentic youth crisis communication patterns.

Objective:

This study aimed to (1) identify the design elements that clinical supervisors, crisis responders, and youth with lived experience consider necessary for a realistic and useful crisis conversation simulator, (2) operationalize these requirements in a redesigned prototype, and (3) evaluate the redesigned simulator with youth with lived experience.

Methods:

We employed Design-Based Implementation Research (DBIR) across two iterations. In Iteration 1, we developed an initial prompt-based simulator grounded in the Kids Help Phone (KHP) clinical model and evaluated it in 4 focus groups with clinical supervisors (n=5), crisis responders (n=6), and youth with lived experience (n=6). Reflexive thematic analysis identified gaps in realism. In Iteration 2, we redesigned the simulator prompt and interface, introducing youth texting conventions and configurable personas and scenarios, and evaluated the revised simulator in 2 focus groups with youth advisory members (n=19), informed by an informal review of de-identified service conversations.

Results:

In Iteration 1, clinical supervisors, crisis responders, and youth converged on three core requirements: authentic youth communication (indirect disclosure, minimization, short-form vernacular, and variable message length and timing), configurable personas and scenarios spanning presentation and acuity, and a supportive, stage-aware feedback component rather than mid-conversation scoring. In Iteration 2, youth judged the redesigned simulator, which introduced youth vernacular and consistent personas, more realistic than the original, while identifying further targets including message chunking, residual over-disclosure, shifting mood, and closer alignment of the assessor with the clinical model.

Conclusions:

A DBIR approach combining clinical, responder, and lived-experience expertise effectively surfaced and refined the design requirements for a youth crisis conversation simulator, with youth identifying realism gaps that a clinically framed design alone would likely have missed. The findings offer transferable guidance for developing generative AI training tools for crisis responders and set the stage for empirical grounding in service data and controlled evaluation of training outcomes.