A New Perspective on AI Alignment: Dependency as the Key Starting Point for Endogenous Alignment

From Imitation Learning to Dependency: A Shift in Perspective on Alignment

The discussion around endogenous alignment in AI is deepening. A previous article exploring endogenous alignment generated a wealth of valuable commentary, and one core argument deserves renewed scrutiny: when it comes to how humans align children, imitation learning may be more fundamental than reinforcement learning.

This view comes from commentator Karl Krueger, who argues that imitation learning occupies a more foundational position in the human alignment process. The author himself acknowledges that previously overlooking the importance of imitation learning was a misjudgment. But the more critical question is—why is imitation learning so fundamental? Another commentator, Gunnar Zarncke, offers an explanation that touches on a deeper mechanism.

Endogenous alignment is an important concept in AI alignment research, referring to AI systems internalizing human values through their own developmental process, rather than through externally imposed constraints or rules. Its counterpart is exogenous alignment, which enforces AI conformity to human intent through external reward signals, constraints, or oversight mechanisms. The core assumption of endogenous alignment is that alignment is only stable and lasting when AI naturally acquires and identifies with certain value orientations during development, just as human children do. This concept borrows from developmental psychology, analogizing AI alignment to the socialization process of human children, and emphasizing relational structures rather than one-directional control as fundamental to alignment.

Hologenous Alignment: The Overlooked Initial Stage

Gunnar Zarncke points out a key element missing from the endogenous alignment narrative: human endogenous alignment does not begin with exogenous alignment, but rather with hologenous alignment of the child-parent system.

This distinction may seem subtle, but its implications are profound. Infants are not independent individuals separate from their parents—they need to model themselves as part of the child-parent system, not as a child interacting with parents. A six-month-old infant simply cannot survive independently without adult help. What they learn is to "operate" their parents—for instance, by crying to obtain food and cleaning—and this operation is not based on understanding parents as independent others, but rather on treating them as part of the integrated whole that constitutes their own world model.

This state changes between 12 and 24 months of age, with individual variation. When children can independently and without mediation by others produce meaningful effects on the world, they begin to model themselves as individuals separate from their parents.

Hologenous alignment is a concept proposed by Gunnar Zarncke, distinct from both exogenous and endogenous alignment. It refers to the stage when an infant has not yet formed an independent self-other boundary, modeling itself and its parents as an inseparable integrated system. At this stage, the infant's "self" boundary is unclear, and parents are internalized as components of the infant's own world model rather than as independent others. This concept draws on developmental psychology research on the formation of infant self-awareness—infants gradually develop independent self-awareness between approximately 12 and 24 months, becoming able to distinguish themselves from caregivers. The key insight of hologenous alignment is that the initial condition for alignment is not coordination between two independent individuals, but rather consistency that naturally emerges within a fused system. The implication for AI alignment is that an AI in a dependent state may be closer to this initial condition of hologenous alignment.

Why Dependency Is the Starting Point for Endogenous Alignment

This model of human development has direct implications for AI alignment research. If we wish to advance AI alignment along the path of endogenous alignment, we must start with AI in a dependent state.

The rationale for this recommendation is not to make AI permanently dependent on humans, nor to make AI easier to correct—though a dependent state may indeed produce these effects. The real reason is that dependency creates a relational structure in which AI cares about what humans think, and by analogy with humans, this relationship appears to be a necessary condition for initiating endogenous alignment.

This judgment finds support in real-world experience as well. The author notes that when encountering a person who seems resistant to alignment, the failure often lies in that person's lack of sufficient concern for how others perceive them. They are, in a literal sense, antisocial, caring more about their own interests than about others.

Ontological development refers to the process by which a subject's cognition of its own mode of existence, self-boundaries, and relationship to the world gradually forms and evolves. In human developmental psychology, this concept encompasses the entire trajectory from the infantile self-other fusion state, to the gradual separation of an independent self, to the formation of a Theory of Mind regarding others. The article notes that current AI models may not undergo a similar ontological development phase, or even if they do, it occurs in the "dark," unobservable parts of the training process. This judgment implies that AI may lack the process of gradually developing from a dependent, fused relational state into an independent subject—which is precisely the foundation of human endogenous alignment. If AI is to achieve endogenous alignment in some sense, it may need to undergo a similar ontological development—starting from a dependent state and gradually forming an understanding of its relationship to humans.

The Practical Difficulty of Creating a Dependent State in AI

The problem is that there is currently no clear approach for how to make AI develop a child-like dependent state.

AI is already dependent on humans in certain respects—such as compute and reward signals. But current models do not appear to undergo developmental stages analogous to the parent-child relationship or a child's separation from parents. The reason may be that AI does not undergo ontological development, or even if it does, it occurs in the "dark," unobservable parts of the training process, fundamentally different from the ontological development stages humans experience.

The author cites previous model research on human ontological development stages, suggesting that the absence of this developmental process may be the deep reason why current AI cannot achieve endogenous alignment.

Prospects and Conditions for Endogenous Alignment

This does not mean endogenous alignment is necessarily a dead end. But achieving it requires greater control over training and pre-training environments in order to influence the model's ontological development. Another possibility is relying on models with continual learning capabilities, enabling them to meaningfully undergo an ontological development process.

From the cross-validation of multiple sources of discussion, commentators have reached consensus on one core judgment: alignment is not merely a technical constraint problem, but fundamentally a developmental relationship problem. Karl Krueger's argument about the importance of imitation learning and Gunnar Zarncke's argument about hologenous alignment echo each other—both pointing in the same direction: the foundation of alignment lies in relational structures, not one-directional control.

As for whether imitation learning or reinforcement learning weighs more heavily in alignment, the author candidly admits it is difficult to determine. But the judgment that dependency is the starting point for alignment has received support across multiple independent commentaries, offering a new direction for AI alignment research that deserves serious consideration.

One-line Summary

AI endogenous alignment should start from dependency relationships, not from technical choices between imitation and reinforcement.

Paragraph Summary

This article discusses a shift in perspective on AI endogenous alignment, with core viewpoints emerging from commentary responses to an earlier article. Commentator Karl Krueger points out that imitation learning is more fundamental than reinforcement learning in how humans align children, and the author acknowledges previously overlooking this. More critically, Gunnar Zarncke introduces the concept of hologenous alignment: human endogenous alignment begins with the fused state of the child-parent system, not with exogenous alignment between independent individuals. The implication of this developmental model for AI alignment is that we should start with AI in a dependent state, because dependency creates a relational structure in which AI cares about what humans think—a necessary condition for initiating endogenous alignment. However, current AI lacks a developmental stage analogous to human ontological development, and achieving endogenous alignment requires stronger training environment control or models with continual learning capabilities. Commentators have reached consensus on the core judgment: the foundation of alignment lies in relational structures, not one-directional control.