Artificial intelligence in pediatric dentistry: reliability and readability analysis of AAPD FAQ responses

Access to accurate and understandable information for parents in pediatric dentistry is critically important. This study aimed to evaluate the information reliability and readability of nine different artificial intelligence (AI) chatbots, referencing the American Academy of Pediatric Dentistry (AAPD) FAQs. Responses to AAPD FAQs from AI chatbots (ChatGPT-4o, ChatGPT-4.1, ChatGPT-5, Gemini 2.5 Pro, Claude 4, Perplexity, Grok 4, DeepSeek V3, LLaMA) were independently evaluated by two experts using the DISCERN questionnaire for information reliability and the Flesch Reading Ease Score (FRES) and Flesch-Kincaid Grade Level (FKGL) metrics for readability. Given the repeated-measures (within-question) design, the Friedman test with post-hoc Wilcoxon signed-rank comparisons (Bonferroni-corrected) was used for statistical analyses. Statistically significant differences ( p < 0.05) were found among AI models in both information reliability (DISCERN) and readability (FRES, FKGL) scores. Perplexity provided the highest information reliability (DISCERN: 74.93 ± 4.79), while Grok 4 yielded the highest readability (FRES: 78.36 ± 11.04); however, Grok 4 also had the lowest reliability score (DISCERN: 46.40 ± 8.48). Claude 4 and DeepSeek V3 demonstrated balanced and above-average performance across all metrics. No AI model achieved the best performance simultaneously across all metrics. AI chatbots hold clear potential for informing parents in pediatric dentistry, but reliability and readability differed across models, with none performing best on both. An apparent readability–reliability trade-off in some models did not reach statistical significance and should be read as descriptive rather than causal. Clinicians should consider each model’s strengths and weaknesses, verify information from independent sources, and position AI as a supportive tool. Models exhibiting balanced performance in the present evaluation, such as Claude 4 and DeepSeek V3, may represent promising options for clinical applications, although model performance may change with future updates.