Comprehensive Summary
The study examined how ChatGPT-4 performs when answering clinical questions about upper tract urothelial carcinoma (UTUC). The authors prompted 77 questions for ChatGPT to answer. 60 questions were from the 2025 European Association of Urology UTUC guidelines, and rest were drawn from frequently asked questions by patients. When two urologists reviewed the responses and scored each answer, 71 out of 77 responses (92.2%) were evaluated as accurate. Specifically, AI’s performance was perfect on all the frequently asked questions. Its performance on questions regarding diagnosis, disease management, and classification were also perfect. However, it did show very low accuracy for follow-up questions. Notably, some of the responses were completely incorrect, raising issues regarding patient safety when used without physician’s support.
Outcomes and Implications
The findings highlight that ChatGPT-4 is capable of providing very highly accurate information, especially for general education on UTUC. However, it is worrisome that the model generated false information when it was asked more complex and follow-up questions. Such patterns are a common limitation observed in language models today. Many models are strong when it comes to broad knowledge, but become vulnerable when prompted to answer complicated questions that require up-to-date information and personalized decision-making. ChatGPT-4 could serve as an effective tool for physicians when educating patients and summarizing general management concepts to better support patients. Regardless, it must not be used without oversight of physicians for making clinical decisions, until AI models are equipped with real-time updates and physician verification.