Causal-Reasoning World Models: Toward AI’s Subjectivity through Backward Generation

Abstract Model interpretability is conventionally equated with “nameable relations”, confining causal modeling to a noun-centric ontology and capping AI’s reasoning at whatever can be named. This conflates two distinct layers of interpretability: 1) capturing relational information, and 2) explaining it through human language, where the latter should never bound the former. Resolving this requires a paradigm shift from conventional noun-based modeling to proposed verb-based modeling, in which AI constructs a world model and reasons within it as an agent, rather than merely recording pre-specified relations as current models do. We develop this verb-based paradigm through conceptual analysis grounded in philosophical foundations, and introduce possible timing distribution , an individual-level construct that goes beyond the i.i.d. assumption underlying traditional statistics. Applied to longitudinal EHR data from 3 , 276 breast cancer patients, the proposed paradigm spontaneously discovers significant patient trajectories and instantiates a What-If Machine for individual-level counterfactual causal deduction. To our knowledge, these are the first demonstrations of their kind, with particular promise for patient-specific precision medicine and rare-disease care. More broadly, this work lays the theoretical groundwork for AI’s subjectivity : not yet its realization, but a necessary first step toward it.