ChatGPT sits the DFPH exam: large language model performance and potential to support public health learning

Abstract

Abstract Background Artificial intelligence-based large language models, like ChatGPT, have been rapidly assessed for both risks and potential in health-related assessment and learning. However, their applications in public health professional exams have not yet been studied. We evaluated the performance of ChatGPT in part of the Faculty of Public Health’s Diplomat exam (DFPH). Methods ChatGPT was provided with a bank of 119 publicly available DFPH question parts from past papers. Its performance was assessed by two active DFPH examiners. The degree of insight and level of understanding apparently displayed by ChatGPT was also assessed. Results ChatGPT passed 3 of 4 papers, surpassing the current pass rate. It performed best on questions relating to research methods. Its answers had a high floor. Examiners identified ChatGPT answers with 73.6% accuracy and human answers with 28.6% accuracy. ChatGPT provided a mean of 3.6 unique insights per question and appeared to demonstrate a required level of learning on 71.4% of occasions. Conclusions Large language models have rapidly increasing potential as a learning tool in public health education. However, their factual fallibility and the difficulty of distinguishing their responses from that of humans pose potential threats to teaching and learning.