Performance of large language models on undergraduate endodontic multiple-choice questions

This study aimed to evaluate the accuracy and consistency of responses provided by three large language models (LLMs), ChatGPT-5.2, Gemini-3, and DeepSeek-V3.2, to multiple-choice questions based on undergraduate endodontic education, asked on different days and at different times of the day. A total of 60 text-based multiple-choice questions were developed across six undergraduate endodontic topics: dental caries, pulpitis, apical periodontitis, periapical abscess, root fracture, and root resorption. Each question was presented to ChatGPT-5.2, Gemini-3, and DeepSeek-V3.2 at three time points per day (morning, afternoon, and evening) over four consecutive days. Accuracy and response consistency were analyzed using SPSS and R software, with statistical significance set at p < 0.05 and a 95% confidence interval. ChatGPT-5.2 and Gemini-3 demonstrated significantly higher accuracy and consistency than DeepSeek-V3.2 ( p < 0.001 and p = 0.004, respectively). Model performance varied according to question category. Accuracy differed significantly across categories for ChatGPT-5.2 and Gemini-3, whereas consistency was influenced by question category only in ChatGPT-5.2. Model performance remained largely stable across different assessment times. Advanced LLMs demonstrated promising performance in answering undergraduate endodontic multiple-choice questions and may serve as useful adjunctive tools in dental education. However, differences among models and variations in performance across topics highlight the need for critical evaluation of AI-generated responses before their educational use.