Introduction: Large language models (LLMs) are increasingly applied in healthcare, yet their diagnostic accuracy in endodontics remains underexplored. This study evaluated the performance of four chatbots—OpenAI GPT-4o, Microsoft Copilot, Google Gemini, and Gemini Advanced (GeminiA)—on endodontic case vignettes. Methods: Seven clinical cases from the American Association of Endodontists newsletter were presented to each chatbot at two time points (June–July 2024). Fifty-six responses were collected and independently scored by two board-certified endodontists. Diagnostic accuracy (pulpal, apical, and overall) was recorded as a binary outcome. Response quality was assessed using a modified Global Quality Score (mGQS, 5-point), reasoning accuracy (6-point), and completeness (3-point). Secondary outcomes included readability (Flesch–Kincaid grade level) and word count. Mixed-effects models evaluated differences among chatbots. Results: Overall diagnostic accuracy was 69.6% (39/56), with significant differences across chatbots (p = 0.006). GeminiA achieved the highest scores across all qualitative measures (mGQS 4.93 ± 0.27; reasoning 5.93 ± 0.27; completeness 3.0 ± 0.00). GPT-4o also demonstrated high performance (mGQS 4.71 ± 0.47; reasoning 5.64 ± 0.50; completeness 2.79 ± 0.43). Copilot consistently underperformed. Readability exceeded college level across chatbots, and word counts varied, with Copilot having the shortest and Gemini having the longest responses. Conclusions: In this exploratory study, advanced LLMs, particularly GeminiA and GPT-4o, outperformed Copilot and Gemini in endodontic diagnosis and reasoning quality. However, these findings should be interpreted with caution, given the limited number of cases and use of publicly available datasets that may have been included in model training. Further validation using larger, de novo case sets is warranted before these tools can be recommended as adjuncts for education or clinical decision support. Clinical significance: Large language model chatbots show promise as adjunctive tools for endodontic diagnosis. Understanding their strengths and limitations may help clinicians and students critically interpret AI-generated recommendations and support clinical decision-making.
Park et al. (Thu,) studied this question.