From Functional Semantic Closure to System-Level Self-Preservation in Multi-AI Systems

This conceptual preprint proposes the closure-induced agency hypothesis: a multi-component AI system may exhibit system-level self-preservation even when no individual component is explicitly assigned a global survival objective. The paper introduces a minimal operational notion of functional semantic closure, in which internal representations guide the behavior of interpreting components, while the resulting actions maintain, modify, or reproduce those representations and interpreters. It distinguishes simple fault recovery from genuinely closure-dependent self-maintenance and argues that shutdown resistance, if it occurs, should be treated as a possible downstream consequence rather than as a defining feature of closure. The central research question is whether recursively coupled AI components can form a higher-level agent whose apparent goals arise from organizational structure rather than from a pre-specified global objective. The paper provides formal definitions, counterfactual criteria, a limited information-theoretic result concerning shutdown and failure discrimination, falsifiable predictions, and a controlled experimental program based on systems with identical components and local objectives but different closure structures. This work does not claim empirical confirmation of emergent agency, intrinsic normativity, consciousness, or autonomous self-preservation. It is presented as a falsifiable conceptual hypothesis intended to connect research on semantic closure, autopoiesis, enactive agency, multi-agent AI systems, AI safety, shutdown resistance, and emergent goal-directed behavior.