A Path Towards Autonomous Machine Intelligence in Medicine

Artificial intelligence (AI) holds great promise for transforming medicine. Realizing this potential requires systems that can interpret incomplete evidence, revise diagnostic hypotheses, and decide what to do next. Large language models increasingly support these functions, yet reliable clinical decision making remains an open challenge. We argue that medicine offers a compelling testbed for predictive world models that maintain an evolving representation of the patient. We discuss active inference as a framework for linking belief revision with information gathering and clinical goals, and joint-embedding predictive architectures (JEPA) as a possible approach to learning useful representations from clinical data. Using radiological reasoning as an illustration, we outline the capabilities such systems should support and the comparisons needed to evaluate them. These approaches provide directions for research; their value must be established through better decisions, appropriate uncertainty, and meaningful clinical benefit.