How Do Deepseek & Chatgpt Perform On Orthodontic Licensing Examinations༟

This study aimed to evaluate the limitations of large language models in orthodontics by comparing their performance on licensing exam questions categorized by knowledge domains and error types. 396 text-based questions from the Chinese National Orthodontic Specialist Licensing Examination were collected and structured through dual taxonomies:(1) "knowledge domains" including foundational biomechanical principles, interdisciplinary medical integration, specialized orthodontic theory, and clinical decision-making skills; (2) "error types" including factual inaccuracies, reasoning deficits, and contextual misinterpretations. Methodological rigor was ensured through context-isolated prompts, temperature-controlled outputs, and specialist-verified response adjudication. McNemar's exact tests with continuity correction were used to analyze performance disparities, adopting a two-tailed α level of 0.05. DS demonstrated significantly higher overall accuracy than GPT (p < 0.05). Subgroup analysis of the four knowledge bases revealed statistically significant differences (p < 0.05) only in the first two domains. While factual inaccuracies were most common in both models, DS demonstrated a higher proportion of reasoning deficits compared to GPT (25.6% vs. 12.6%), with a significant difference in error distribution confirmed by chi-square analysis (χ² = 39.5, p < 0.01). The study shows that despite DS's overall superior performance in the orthodontic licensing examination compared to GPT, it has not yet overtaken GPT in specialized fields. The two types of models exhibit a complementary situation in the orthodontic professional context: open-source models have an edge in systematic knowledge retrieval, while closed-source models demonstrate greater adaptability in unstructured clinical reasoning.