Comparing large language models and search engine responses to common orthodontic questions

Abstract

Background Large Language Models (LLMs) highlight their potential in supporting patient education and self-management. Their performance in responses to orthodontic questions has yet to be explored. Objectives This study aims to compare the quality, empathy, readability, and satisfaction of responses from LLMs and search engines on common orthodontic questions. Methods Forty-five common orthodontic questions (six categories) and a prompt were developed, and a self-designed multidimensional evaluation questionnaire was constructed. Questions were presented to 5 LLMs and 3 search engines on December,22,2024. The primary outcomes were the median expert-rated scores of LLMs versus search engine responses on quality, empathy, readability, and satisfaction, using 5- or 10-point Likert scales. Results LLMs scored significantly higher than search engines in quality (4.00 vs. 3.50, p<0.001), empathy (3.75 vs. 3.50, p<0.001), readability (4.00 vs. 3.75, p<0.001), and satisfaction (8.00 vs. 7.25, p<0.001). LLM-generated responses were rated significantly higher than those from search engines in therapeutic outcomes category, appliance selection category and cost category. Conclusions In this cross-sectional study, the LLMs, particularly GPT-4o, outperformed search engines. These results indicate the potential of LLMs as supplementary tools for orthodontic patient education and self-management.