Development of a generative AI agent for family support in implementing family-based treatment for children and adolescents with anorexia nervosa

Editorial reading guide · Trust

What evidence would a specialized AI family-support tool need before I relied on it?

The retrieval-based, safety-gated system produced responses judged appropriate in 91.6% of queries, but important information and escalation failures remained. Its safeguards make accountability discussable without demonstrating that families can safely rely on it.

Study context

Six pediatric mental-health clinicians wrote 477 simulated family questions and each evaluated the responses to their own questions from a specialized family-based-treatment agent.

Age / stage: Children and adolescents with anorexia nervosa; families not enrolled

Read with three questions

1. How do the knowledge base and escalation rules constrain responses?
2. What do the remaining information and escalation errors reveal?
3. Why is clinician self-rating of simulated queries insufficient to establish family benefit?

Keep the limits in view

This single-center internal evaluation used no independent cross-rating and included reviewers partly aware of design aims. No actual families tested usability, and caregiver burden, adherence, sustained dialogue safety, and child outcomes were not measured.