World Foundation Models and Agentic AI in Medicine: Physiological Simulation, Digital Twins, and Clinical Decision-Making

Many foundation models that have entered medicine are language models, although foundation models for imaging and other modalities have also emerged. A second class is now emerging from robotics and autonomous driving, trained not to produce text but to predict how an environment evolves under action, and marketed under the designation world foundation model. We argue that medicine should approach this class with a specific form of caution. Medicine already possesses world models in the mechanistic simulators and digital twins built from physiology and physics, and the constraint on those has never been the volume of available data. The constraint is identifiability. A simulator learned from observational clinical records inherits the treatment policies that generated them, and rolling such a model forward under an alternative treatment recovers the practice pattern rather than the underlying physiology. We distinguish the settings where learned world models are likely to succeed from those where they will reproduce confounding at scale, and set out what validation would require.