Can ChatGPT-4o really pass medical science exams? A pragmatic analysis using novel questions

Abstract

Abstract ChatGPT apparently shows excellent performance on high level professional exams such as those involved in medical assessment and licensing. This has raised concerns that ChatGPT could be used for academic misconduct, especially in unproctored online exams. However, ChatGPT has also shown weaker performance on questions with pictures, and there have been concerns that ChatGPT’s performance may be artificially inflated by the public nature of the sample questions tested, meaning they likely formed part of the training materials for ChatGPT. This led to suggestions that cheating could be mitigated by using novel questions for every sitting of an exam and making extensive use of picture-based questions. These approaches remain untested. Here we tested the performance of ChatGPT-4o on existing medical licensing exams in the UK and USA, and on novel questions based on those exams. ChatGPT-4o scored 94% on the United Kingdom Medical Licensing Exam Applied Knowledge Test, and 89.9% on the United States Medical Licensing Exam Step 1. Performance was not diminished when the questions were rewritten into novel versions, or on completely novel questions which were not based on any existing questions. ChatGPT did show a slightly reduced performance on questions containing images, particularly when the answer options were added to an image as text labels. These data demonstrate that the performance of ChatGPT continues to improve and that online unproctored exams are an invalid form of assessment of the foundational knowledge needed for higher order learning.

Ähnliche Arbeiten