Evaluation of large language model responses to different types of questions based on European Society of Endodontology position statements

Objectives: This study aimed to compare the performance of DeepSeek-V3, ChatGPT 4.0, ChatGPT 4.5, Gemini Advanced 2.0, and Grok in answering single-correct answer multiple-choice (SCQ), multiple-correct answer multiple-choice (MCQ), binary (true/false) (BQ), and open-ended questions (OEQ) based on the European Society of Endodontology (ESE) position statements.Materials and Methods: A total of 70 questions (24 SCQ/MCQ, 29 BQ, and 17 OEQ) were posed to five LLMs. Responses were scored on a 0–10 scale for comprehensiveness, scientific accuracy, clarity, and relevance. SCQ, MCQ, and BQ formats were additionally evaluated for correctness (1 = correct, 0 = incorrect). Data were analyzed using Pearson’s chi-square test and the Kruskal-Wallis H test (p < 0.05).Results: In the SCQ/MCQ format, Gemini Advanced 2.0 scored significantly higher than Grok. In the BQ format, ChatGPT 4.5 and Gemini Advanced 2.0 achieved significantly higher scores than ChatGPT 4.0 and Grok. In the OEQ format, ChatGPT 4.5 obtained significantly higher scores than Grok (p < 0.05). Accuracy rates were 67.9% for ChatGPT 4.5, 66.0% for Gemini Advanced 2.0, 58.5% for DeepSeek-V3, 52.8% for ChatGPT 4.0, and 49.1% for Grok (p > 0.05). In all models, scientific accuracy scores were significantly lower than relevance scores (p < 0.05).Conclusions: ChatGPT 4.5 and Gemini Advanced 2.0 demonstrated stronger performance than the other models, whereas Grok showed lower response-content performance. However, since LLMs may appear convincing even when generating incorrect clinical information, they should be used only as supplementary tools in endodontic education and preliminary information retrieval. Clinically relevant outputs should be verified by an endodontist or an appropriately qualified clinician.