Abstract
Modern AI expands the modeling of economic systems by enriching the representation of agents, environments, information, and interaction. I introduce Economic World Models (EWMs) as a model class in which agents, environments, and their interactions may be theory-specified, learned from data, or hybrids of the two. I define Data-Driven Generative Equilibrium (DDGE), the equilibrium concept for data-driven EWMs in which behavior, beliefs, generated data, and learned system components must be jointly consistent. DDGE operationalizes the Lucas critique when some structural primitives are learned rather than fully specified: a counterfactual may change behavior, behavior may change the training data, and retraining may change the environment being evaluated. I show that frozen simulation and DDGE generally differ, quantify the resulting center-displacement welfare loss through a one-retraining-step residual, identify learning-generated multiplicity and local retraining instability, and separate robust control around a model from equilibrium re-centering of the model itself. Applications to portfolio choice, corporate decision support, AI-mediated lending, and simulated learned economies illustrate how EWMs extend economics from estimating isolated mechanisms to analyzing learned economic systems while preserving equilibrium discipline.