Theory as a Model-Generating System: Rethinking Scientific Reasoning in the Age of Generative Artificial Intelligence

Generative AI makes it cheap to produce many formally coherent scientific models—equations, simulations, and structural representations—shifting bottlenecks from model construction toward model selection and evaluation. This paper proposes redefining theory as a model-generating system and introduces five selection principles: observability, identifiability, intervenability, decision relevance, and complexity control. A worked miniature example illustrates principled model collapse in epidemiological (SEIR-type) modeling.