From chatbots to chairside: how artificial intelligence understands molar incisor hypomineralization in pediatric dentistry

Large language models (LLMs) are increasingly integrated into clinical workflows, yet their utility in pediatric dentistry remains unclear. This study evaluated and compared the clinical quality, readability, and specificity of responses generated by four contemporary LLMs to questions regarding Molar Incisor Hypomineralization (MIH) from dentists and parents. A standardized set of 50 open-ended and true/false MIH questions was administered to ChatGPT-4o, ChatGPT-5, Gemini 2.0 Flash, and DeepSeek-V3.1. Two independent pediatric dentists evaluated response quality using the Modified Global Quality Scale (MGQS), and the mean score served as the primary outcome. Readability was assessed using the Flesch Reading Ease Score (FRES) and the Flesch–Kincaid Grade Level (FKGL). Data were analyzed using a repeated-measures design: MGQS scores were compared via the Friedman test, and readability metrics were compared using repeated-measures ANOVA or the Friedman test based on normality (Shapiro–Wilk test); multiple comparisons were evaluated using the Bonferroni-corrected Conover test. Inter-rater agreement across all 200 responses was determined using unweighted Cohen’s kappa (κ = 0.662; 95% CI, 0.565–0.758). Significant differences in MGQS scores persisted among the models for open-ended questions directed to both parents (Friedman χ²(3) = 32.52, p < 0.001) and pediatric dentists (χ²(3) = 18.03, p < 0.001); DeepSeek-V3.1 scored significantly higher than the other three models in both subgroups (Bonferroni-corrected Conover test, all p < 0.05), while ChatGPT-5, Gemini 2.0 Flash, and ChatGPT-4o did not differ significantly from one another in most pairwise comparisons. No significant differences were found for true/false questions (both subgroups p > 0.10). Readability differed significantly among models in every subgroup: Gemini 2.0 Flash and DeepSeek-V3.1 generally produced more accessible content (higher FRES, lower FKGL) than ChatGPT-5 and ChatGPT-4o, although the least-readable model varied by subgroup rather than being uniformly attributable to a single model. DeepSeek-V3.1 demonstrates a clear clinical quality advantage for open-ended queries in pediatric dentistry, while Gemini 2.0 Flash and DeepSeek-V3.1 provide more readable text overall. LLMs exhibit heterogeneous, subgroup-dependent performance and should be used with caution as adjunctive tools in MIH-related education and clinical care. Readability profiles should not be broadly generalized across different question formats.