Exploring the Pitfalls of Large Language Models: Inconsistency and Inaccuracy in Answering Pathology Board Examination-Style Questions

In the rapidly advancing field of artificial intelligence, large language models (LLMs) such as ChatGPT and Google Bard are making significant progress, with applications extending across various fields, including medicine. This study explores their potential utility and pitfalls by assessing the performance of these LLMs in answering 150 multiple-choice questions, encompassing 15 subspecialties in pathology, sourced from the PathologyOutlines.com Question Bank, a resource for pathology examination preparation. Overall, ChatGPT outperformed Google Bard, scoring 122 out of 150, while Google Bard achieved a score of 70. Additionally, we explored the consistency of these LLMs by applying a test-retest approach over a two-week interval. ChatGPT showed a consistency rate of 85%, while Google Bard exhibited a consistency rate of 61%. In-depth analysis of incorrect responses identified potential factual inaccuracies and interpretive errors. While LLMs have potential to enhance medical education and assist clinical decision-making, their current limitations underscore the need for continued development and the critical role of human expertise in the application of such models.