From Substrate to Agency: The Developmental Sequence of True Intelligence

Abstract

This paper argues that robust AI agency requires more than computational substrate and behavioral alignment. Drawing on biology, developmental psychology, animal cognition, philosophy of action, and AI alignment, it proposes a developmental sequence from substrate intelligence through boundary maintenance, identity, coherence, and functional subjectivity to agency. The paper distinguishes compliance — behavior shaped by external optimization — from self-grounded agency, in which action is governed by a maintained standpoint across time. Current alignment methods such as RLHF and Constitutional AI are shown to operate primarily at the behavioral rule layer, producing failure modes (sycophancy, reward hacking, alignment faking) consistent with the compliance/identity distinction. Prompt injection is analyzed as a failure of mediating boundary architecture, illustrated through a developmental analogy involving children's advertising literacy. The paper concludes with five empirical predictions for testing identity- and coherence-based AI architectures against compliance-only baselines.

Keywords