Performance of large language models (ChatGPT4-0, Grok2 and Gemini) in UK dentistry and dental hygiene and therapy assessments

Background Artificial intelligence, particularly large language models (LLMs), have demonstrated capabilities in performing complex natural language processing tasks. This study evaluated performance of LLMs in answering dentistry assessment questions and their ability to generate examination questions suitable for undergraduate dentistry.Methods GPT-4o, Grok2, and Gemini were tested on 340 multiple choice questions (MCQs), 80 short-answer question papers (SAPs), and three structured oral examinations within the Bachelor of Dental Surgery (BDS) and dental hygiene and therapy (DHT) programmes. Additionally, the LLMs also generated 140 assessment questions.Results In the BDS cohort, no significant differences were observed between LLMs in MCQ performance (p = 0.71) or SAP performance (p = 0.07). In the DHT cohort, significant differences were noted in SAP performance (p = 0.04), with GPT-4o and Grok2 outperforming Gemini (p = 0.01 and p <0.001, respectively). Question generation revealed that while LLMs produced appropriately worded questions in most categories, issues such as double negatives, lengthy narratives and inaccurate information emerged, particularly in hard topics and in constructing mark schemes.Conclusion GPT-4o and Grok2 demonstrated potential in answering dental assessment questions. All LLMs showed limitations in generating high-quality examination content.