Appropriateness of Recommendations Provided by ChatGPT to Interventional Radiologists
Maxime Barat, Philippe Soyer, Anthony Dohan
Abstract
Purpose: The purpose of this study was to analyze the answers provided by ChatGPT to various questions in the field of interventional radiology (IR) and compare their correctness to a consensus of three interventional radiologists. Materials and Methods. Twenty questions were proposed to the online chat based ChatGPT about interventional radiology as indications for exams, complications and technical points. Answers were independently evaluated by 3 experts in interventional radiology with 10, 15 and 30 years or experiences and classified as correct, almost correct, or wrong. Results: Considering the consensus between the two IR, nine (9/20; 45%) were considered wrong, three (3/20; 15%) were considered almost correct and eight (8/20; 40%) were considered correct. Conclusion: ChatGPT can be useful in IR but the users must be aware of limitations such as too vague or inaccurate answers for the majority of questions.