Performance of artificial intelligence chatbots in restorative dentistry: a comparative analysis of clinical and theoretical specialty examination questions

Large language model (LLM)-based artificial intelligence (AI) systems are used in healthcare education, but their performance on specialty-level restorative dentistry questions remains unclear. This study aimed to compare the performance of six AI systems on Restorative Dentistry questions from the Turkish Dentistry Specialization Examination (DUS). A total of 195 text-based questions from DUS between 2012 and 2025 were included in this cross-sectional observational study and classifiedas theoretical or clinically oriented. GPT-5.2 Instant, Claude Sonnet 4.5, Microsoft Copilot, Gemini 3 Flash, Perplexity, and DeepSeek-V3.2 were evaluated. Responses were compared with official answer keys. Cochran’s Q test, Holm-adjusted McNemar tests, and mixed-effects logistic regression was used for statistical analyses. Accuracy differed significantly among AI systems ( p < 0.001). Overall accuracy ranged from 83.6% for Perplexity to 96.4% for Gemini 3 Flash, with GPT-5.2 Instant, Claude Sonnet 4.5, Microsoft Copilot, and DeepSeek-V3.2 achieving 90.3%, 87.7%, 89.7%, and 90.8%, respectively. After Holm adjustment, Gemini 3 Flash significantly outperformed Perplexity, Claude Sonnet 4.5, and Microsoft Copilot. In mixed-effects analysis, Gemini 3 Flash had lower odds of an incorrect response than GPT-5.2 Instant (OR = 0.185, 95% CI: 0.059–0.581; p = 0.004), whereas Perplexity had higher odds (OR = 2.683, 95% CI: 1.219–5.909; p = 0.014). Neither question type nor subtopic showed a significant overall effect. Contemporary AI systems demonstrated high accuracy on specialty-level dental examination questions, with significant performance differences among systems. These findings support their potential as complementary resources for dental education and examination preparation, but do not establish educational effectiveness, reasoning ability, or clinical safety.