The Golden Cage: How Algorithmic Nurturing Systemically Erodes Human Autonomy Under Artificial General Intelligence

The proliferation of Artificial General Intelligence (AGI) presents a systemic paradox within complex socio-technical systems: while enhancing efficiency, AGI may subvert human autonomy through comfort-based alignment rather than overt coercion. Although algorithmic management research has theorized surveillance-based control—the “Iron Cage”—it has largely overlooked the voluntary erosion of autonomy driven by algorithmic nurturing and its recursive feedback loops. This study proposes the Algorithmic Nurturing Perspective (ANP) by synthesizing five foundational behavioral theories—bounded rationality, expectancy theory, prospect theory, goal-setting theory, and social information processing theory—reinterpreted within the AGI context. ANP elucidates how AGI systemically reconfigures socio-technical systems through three self-reinforcing causal feedback loops: the automation of choice, the externalization of emotional regulation, and isolation from social reality. We derive eight propositions illustrating how these mechanisms trigger individual-level cognitive dependence and systemic regression, which subsequently emerge as organizational-level pathologies, including learning myopia, leadership degeneration, and declining strategic decision quality. By introducing the “Golden Cage” as a novel conceptual lens, the ANP shifts the AI governance discourse from an “efficiency-vs-coercion” framework to an “autonomy-vs-compliance” paradigm. We further propose Cognitive Friction Design as a multilevel balancing intervention to mitigate these risks. As a self-contained theoretical contribution, this framework stands independently of any single empirical test while remaining falsifiable: a mechanism-differentiated system dynamics simulation demonstrates the internal coherence of the proposed feedback structure, and proposition-by-proposition falsification conditions specify the empirical pathway for causal verification in future research. We further identify balancing-dominant conditions under which algorithmic assistance may enhance, rather than erode, human autonomy.