The Limits of Predictive Intelligence: World Models, World-Binding, and Responsibility

Contemporary approaches to artificial intelligence increasingly assume that powerful World Models could form the foundation of future Artificial General Intelligence (AGI). The underlying assumption is that a system's ability to model relationships, predict future states, and generate increasingly comprehensive representations of its environment simultaneously constitutes an approximation to understanding the world. The present paper fundamentally challenges this assumption and argues that world modeling and world-binding are ontologically different processes. Drawing on the concepts of Weltbindung, Eigenzeit, and non-closure, the paper argues that the world must not be confused with its representation. It follows that the central limits of World Models do not primarily lie at the level of modeling, but at the level of decision and responsibility. A World Model can describe relationships, recognize patterns, and simulate consequences. However, it cannot generate the conditions under which the world becomes binding for a system in the first place. Weltbindung does not arise through the organization of information, but through irreversible differentiation, real consequences, and embeddedness in an irreversible history that cannot be completely calculated. The paper develops the thesis that World Models implicitly rely on a globalization of relevance. They attempt to translate local, historically developed, and embodied forms of world-relation into universal model structures. In doing so, however, they lose access to the conditions under which relevance arises in the first place. Relevance does not appear as a property of information, but as the consequence of concrete Weltbindungen within an Eigenzeit. On this basis, it is argued that the central difficulty of future AGI systems may not lie in insufficient computational power, missing data, or incomplete models. Rather, the actual limit may consist in the fact that Eigenzeit, irreversibility, and Weltbindung cannot be transformed into transferable model structures without loss. The description of a bond is not identical with the bond itself. The simulation of a consequence is not identical with the consequence. The representation of a history is not identical with its emergence. Finally, the paper argues that responsibility does not emerge from maximal predictive capability, but from dealing with persistent uncertainty. If the world is grounded in non-closure, then not-knowing is not a deficiency of intelligence but a condition of responsible action. From this perspective, the limits of World Models do not appear as technical deficits, but as the expression of a fundamental difference between model and world that extends beyond previous critical map–reality paradigms.