Generating occlusion clinical cases with large language models: a comparative evaluation of claude, gemini, and qwen in undergraduate dental education

Background Large language models (LLMs) are increasingly being explored for dental education, but their ability to generate undergraduate occlusion cases has received limited attention. This study compared the quality of cases generated by Claude Sonnet 5, Gemini 3.6 Flash, and Qwen 3.8-Max using a standardized clinical reference framework. Materials and methods The models were accessed through their web-based chat interfaces with default settings (1–15 August 2026). Each model generated 54 cases using the same reference framework and prompt, yielding 162 cases. Two blinded prosthodontic experts independently assessed each case across six domains (Clinical Realism, Clinical Accuracy, Completeness, Educational Value, Guideline Adherence, and Difficulty Level) on 1–5 scales (total score 6–30). Evaluator scores were averaged for between-model comparisons. Differences were analyzed using Kruskal–Wallis tests and Bonferroni-adjusted pairwise Mann–Whitney U tests. Inter-rater reliability was assessed using an intraclass correlation coefficient (ICC). Results Claude achieved the highest mean total score (27.94 ± 0.99), followed by Gemini (27.41 ± 1.39) and Qwen (27.13 ± 1.49). Overall scores differed significantly among the models (H = 8.910, p = 0.012; ε 2 = 0.043). Claude scored significantly higher than Qwen (adjusted p = 0.006; r = 0.296), whereas the other pairwise comparisons were not significant. Overall differences were also found across all six evaluation domains, although pairwise results varied; for example, Gemini scored highest for Completeness and Qwen lowest for Difficulty Level. Inter-rater reliability was poor (single-measure ICC = 0.303; average-measure ICC = 0.465). Conclusion All three models generated highly rated occlusion cases, although their performance varied across evaluation domains. Claude scored significantly higher than Qwen, but the difference was small, and poor inter-rater reliability limits confidence in the comparative findings, which apply only to the tested versions and interface conditions. LLMs may be useful for preparing supplementary teaching materials, provided that cases are reviewed by experts before incorporation into teaching curricula. Further research is needed to assess the consistency of generated cases and determine whether they improve student learning.