Related Experiment Video
Updated: May 12, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Quality of artificial intelligence-generated responses on pediatric celiac disease: Comparative assessment of Open AI
Ashwin Agrawal1,2, Miri Goodman2, Alexander Schosheim1,2
1Division of Pediatric Gastroenterology, Unterberg Children's Hospital at Monmouth Medical Center RWJBarnabas Health Long Branch New Jersey USA.
Objectives:
Many individuals use artificial intelligence tools such as Open AI's ChatGPT and Google's Gemini to assess their symptoms or learn about a disease. However, studies to evaluate responses provided by artificial intelligence tools are lacking. This study aims to evaluate and compare ChatGPT and Gemini in answering pediatric celiac disease-specific questions in terms of performance across accuracy, safety, completeness, and understandability.
Methods:
We developed 34 questions that mimicked questions that parents ask about pediatric celiac disease in four categories: disease symptoms and definition, diagnosis, gluten challenge, and management. Questions were posed to ChatGPT 5.0 and Gemini Pro 2.5. Responses were compared with North American Society for Pediatric Gastroenterology, Hepatology, and Nutrition (NASPGHAN) and European Society for Pediatric Gastroenterology, Hepatology, and Nutrition (ESPGHAN) guidelines, or clinical best practices when guidelines were absent. Three board certified Pediatric Gastroenterologists rated each response on a 5-point Likert scale across four domains: accuracy, safety, completeness, and understandability. Word count and Flesch-Kincaid grade level (FKGL) were recorded.
Results:
Average scores were ≥4.5 for ChatGPT and ≥4.3 for Gemini across all four domains. Gemini responses were significantly longer and had higher FKGL. Interclass correlation coefficients showed good agreement for all rating domains except for Gemini safety ratings and understandability ratings for both models.
Conclusions:
ChatGPT and Gemini both generated highly accurate, safe, complete, and understandable information for common questions about celiac disease. Performance decreased when guidelines were lacking or there was clinical ambiguity. However, readability scores (FKGL) for both models exceeded standards for patient education recommendations for sixth to eighth grade reading level or below suggesting there may be barrier to understandability for the general population.
