Oncologic clinical decision-making is full of nuance. What’s not full of nuance? NCCN treatment algorithms, medical textbooks, and lectures. Enter ChatGPT-4-turbo, to which you can ask highly-individualized clinical questions in niche fields and get immediate thesis-level answers. Sounds too good to be true? Probably for now, but maybe not for long. This study assessed 7 large language models’ performances on the 2021 American College of Radiology (ACR) Radiation Oncology In-Training Examination (TXIT), and ChatGPT-4-turbo led the field with 74.2% correct answers. That’s far from 100% but is comparable to PGY-4 and PGY-5 residents. Lastly, ChatGPT-4-turbo did worse on clinical questions but its score was buffered by stellar performance on stats questions, which makes sense as it has had no rad onc-specific training. TBL: Upper level rad onc residents perform as well on the in-service exam as AI with access to one petabyte of data. | Thaker, AI in Precision Oncology 2024

