Freely accessible large language models for parent education in pediatric immune thrombocytopenia: an expert-rated cross-sectional study of safety, readability, and guideline concordance

Editorial reading guide · Trust

Does the best overall score identify the safest AI explanation for a parent?

The highest composite score chiefly reflected communication quality, while safety-sensitive domains were broadly similar. Two unsafe or hallucinated responses occurred in the highest-scoring system, illustrating why a favorable average cannot settle safety.

Study context

Three blinded clinicians rated 120 answers to forty standardized parent questions generated once by three freely accessible LLMs.

Age / stage: Childhood immune thrombocytopenia; no enrolled parent users

Read with three questions

1. Which domains drove the composite-score advantage?
2. What can two observed failures and zero in other tools tell us about rare risks?
3. How would repeated generations or actual caregiver testing change the evidence?

Keep the limits in view

Each question was asked once and the clinician-created scoring framework was not externally validated or co-developed with caregivers. Rare-event counts and wide intervals cannot establish stable comparative safety or exclude failures in systems with no observed events.