Mehmet Adıgüzel
BACKGROUND This study aimed to evaluate the accuracy, confidence, and overconfidence behavior of large language models (LLMs) in endodontics using questions derived from a national specialty entrance examination. MATERIAL AND METHODS A total of 123 text-based endodontic questions from the Turkish Dental Specialty Examination (2017-2026) were included after excluding annulled and image-based questions. Three LLMs (ChatGPT, Claude, and Gemini) were assessed. Each model answered all questions using a standardized prompt and provided a confidence score (0%-100%). Accuracy was recorded as correct/incorrect. Overconfidence was defined as incorrect responses with ≥90% confidence. Statistical analyses were performed using Cochran's Q test, Friedman test, and post hoc pairwise comparisons with Bonferroni correction. RESULTS Accuracy rates were 89.4% for ChatGPT, 76.4% for Claude, and 90.2% for Gemini, with significant differences among models (χ²(2)=21.00, P<0.001). Confidence scores differed significantly (c2(2)=213.40, P<0.001), with Gemini demonstrating highest confidence (99.51±1.49), followed by ChatGPT (90.88±9.19) and Claude (80.22±11.10). Overconfidence rates were 9.8% for Gemini and 5.7% for ChatGPT, while no overconfident responses were observed for Claude (χ²(2)=14.53, P=0.001). Despite similar accuracy between ChatGPT and Gemini, confidence patterns differed markedly, demonstrating that comparable accuracy does not necessarily reflect comparable reliability. CONCLUSIONS LLMs demonstrated high accuracy in answering text-based endodontic examination questions; however, significant differences were observed in confidence behavior and overconfidence patterns. The presence of high-confidence incorrect responses suggests that accuracy alone may be insufficient to fully evaluate model reliability. These findings highlight the importance of considering confidence-related behavior alongside accuracy when assessing LLM performance in examination-style endodontic tasks.