Doing good science in the age of large language models: CARE framework for radiology AI

This is the authors' abstract. We don't add key points for this paper.

Radiology has always been an early adopter of technology, but progress in artificial intelligence has often outpaced meaningful clinical impact. With the arrival of large language models (LLMs), the field faces both a new frontier and a familiar challenge: turning technical capability into practical value. We argue that doing good science in the LLM era means returning to the fundamentals by asking the right questions, building with clinicians, and focusing on what truly improves care. The proposed CARE framework, which emphasizes Clinical relevance, Appropriate task fit, Real-world feasibility, and Evaluation design, offers a way to keep radiology AI grounded, rigorous, and ultimately practical where it matters most, in patient care. But also moving forward in a way that is economically sustainable and broadly applicable across AI modalities. • AI radiology research should shift from model-driven to problem-driven, grounded in clinical need and reproducibility. • The CARE framework guides radiology AI projects through relevance, fit, feasibility, and evaluation. • Implementation science ensures radiology AI tools are usable, scalable, and impactful in real settings.