The renaissance of information retrieval and the limitations of artificial intelligence

This paper critically examines the limitations of generative AI in medicine and challenges the popular narrative that multimodality alone can rescue its current shortcomings. It argues for a shift toward evidence-based AI grounded in structured, high-quality clinical data rather than unsupported generation. Retrieval-Augmented Generation (RAG) is presented as a promising framework for improving grounding, attribution, and explainability, though its clinical use remains underdeveloped. The paper also highlights the risks of overreliance on commercial cloud infrastructure, particularly with respect to cost, dependency, and research autonomy. It calls for foundational research in multimodal retrieval and specialized models to enable AI systems that are safe, explainable, clinically grounded, and genuinely useful in medicine.