Letter: Performance of ChatGPT and GPT-4 on Neurosurgery Written Board Examinations

Abstract

To the Editor: We appreciate the study by Ali et al1 on the performance of ChatGPT and GPT-4 in neurosurgery written board examinations, and we also appreciate the letter by Zhu and Kong,2 particularly for highlighting the importance of integrating visual capabilities into large language models (LLMs). Recently, medical researchers have been testing ChatGPT's performance in examinations across various countries3-6 and specialties.7-11 As social science researchers studying medical artificial intelligence (AI), we find this enthusiasm among medical researchers noteworthy. Unlike the more conservative views often held by scholars in the humanities and social sciences, medical researchers seem to lean toward technological accelerationism by focusing on LLMs' capabilities. To foster interdisciplinary discussion, we suggest stepping back to examine the broader philosophical and social implications of ChatGPT passing medical examinations. What does this mean? ChatGPT passing medical examinations indicates that natural language processing and machine learning have reached a new height. Specifically, the use of transformer architecture and large-scale pretraining techniques has enabled these models to excel in understanding and generating complex texts. This signifies not only an improvement in model capabilities but also the potential applications of LLMs in specialized fields. However, although ChatGPT (with most studies focusing on this particular popular product) shows such potential, we should not blindly trust it for direct application in professional medical fields. The current ChatGPT is trained on unverified public data sets, and it is not yet suitable for use in specialized medical areas.12 Developing specialized medical LLMs is the right path to truly unlock this potential.13 Future efforts should focus on training and optimizing these models with specific medical data and needs, ensuring their reliability and safety in clinical applications through rigorous validation and control, thus providing valuable support to medical practice.14 What does this not mean? Although ChatGPT can pass medical examinations in some scenarios, it does not mean it has true understanding. From a semantic perspective, understanding involves a deep grasp of concepts and their relationships, whereas LLMs only perform syntactic pattern recognition. This phenomenon can be explained by the Chinese Room Argument, which suggests that the system merely simulates understanding through symbol manipulation without actual comprehension.15 LLMs rely on statistical pattern matching and probability prediction, so they may fail when dealing with problems beyond their training context or involving complex reasoning. Their dependency on context and knowledge limits their applicability in complex and variable situations. Although studies like Ali et al's have found that GPT-4's accuracy is significantly higher than GPT-3.5,1 this only shows an improvement in generating seemingly coherent text, not in understanding. Even if GPT-4 and its successors (GPT-5, GPT-6, etc.) achieve 100% accuracy, it does not mean they possess human intelligence. Therefore, as more researchers test whether ChatGPT can pass medical examinations and as this becomes a hot research topic, we must be wary of the technological accelerationism and the belief in the omnipotence of LLMs underlying such research questions. We cannot expect that LLMs will one day replace human doctors in clinical diagnosis and treatment; ChatGPT's role can and should always be only as an assistant. Does this mean that ChatGPT passing medical examinations is meaningless? In fact, this type of research reveals the great potential and safety of ChatGPT and other LLMs in Medical Education.16 LLMs can be used for examination training and knowledge Q&A for medical students and residents, providing instant feedback and detailed explanations. This can greatly improve learning efficiency and help students better grasp complex medical concepts and clinical skills through personalized tutoring and simulated practice. In this way, LLMs can significantly enhance the quality of medical education and provide strong support for training future doctors. Most importantly, this application improves teaching effectiveness while avoiding the risks and uncertainties associated with direct clinical application. After all, compared with AI applications in other fields (such as education and entertainment), the medical field requires the highest level of control over AI. In this sense, better safe than sorry. Because this is an interdisciplinary perspective, we look forward to any criticism and feedback from the authors and other researchers.

Ähnliche Arbeiten