Can a large language model create acceptable dental board-style examination questions? A cross-sectional prospective...

Numerous studies have shown that large language models (LLMs) can score above the passing grade on various board examinations. Therefore, this study aimed to evaluate national dental board-style examination questions created by an LLM versus those created by human experts using item analysis.

This study was conducted in June 2024 and included senior dental students (n = 30) who participated voluntarily. An LLM, ChatGPT 4o, was used to generate 44 national dental board-style examination questions based on textbook content. Twenty questions for the LLM set were randomly selected after removing false questions. Two experts created another set of 20 questions based on the same content and in the same style as the LLM. Participating students simultaneously answered a total of 40 questions divided into two sets using Google Forms in the classroom. The responses were analyzed to assess difficulty, discrimination index, and distractor efficiency. Statistical comparisons were performed using the Wilcoxon signed rank test or linear-by-linear association test, with a confidence level of 95%.

The response rate was 100%. The median difficulty indices of the LLM and human set were 55.00% and 50.00%, both within the range of "excellent" range. The median discrimination indices were 0.29 for the LLM set and 0.14 for the human set. Both sets had a median distractor efficiency of 80.00%. The differences in all criteria were not statistically significant (P > 0.050).

The LLM can create national board-style examination questions of equivalent quality to those created by human experts.