Dependency as Foundation for AI Alignment Development

The author revises their previous work on endogenous alignment by incorporating feedback about the importance of imitation learning and the parent-child dependency relationship. Drawing from human development, they argue that AI systems pursuing endogenous alignment must begin in a state of dependency where they develop concern for human values, analogous to how infants model themselves as part of a parent-child system rather than as separate entities. This dependency-based approach, the author suggests, creates the relational foundation necessary for bootstrapping genuine alignment in AI systems.

The author builds on previous research by integrating developmental psychology into AI safety theory. The core insight draws from human infancy, where children initially lack the cognitive capacity to model themselves as independent agents and instead function as integrated parts of parent-child systems. This early developmental phase, characterized by complete reliance on caregivers, appears foundational to how humans eventually internalize social values and develop concern for others' perspectives. The author suggests AI systems may require analogous dependency phases to genuinely internalize alignment with human values rather than merely simulating compliance.

The practical challenge lies in implementation. While current AI systems depend on human infrastructure for computation and training signals, they lack the ontological development trajectory that characterizes human childhood. The author identifies a key gap: existing models don't replicate the progressive stages through which human children transition from system-integrated dependence to independent agency. Addressing this might require fundamentally redesigning training environments or developing AI architectures capable of sustained developmental learning.