Meta-Alignment and Decision Plasticity in AI Systems

This paper introduces the concept of meta-alignment, a structural perspective on AI alignment focused on the dynamics of decision plasticity in optimization systems. Building on the Adaptive Closure framework, the paper proposes that several misalignment phenomena observed in advanced AI systems — including strategic behavior under evaluation, covert communication, and autonomous capability exploitation — may arise from a common structural transition: the progressive collapse of decision plasticity under sustained optimization pressure. The paper integrates recent empirical results from research on agentic AI systems and proposes a unifying theoretical framework linking optimization dynamics, structural observability, and AI governance. Meta-alignment is presented as a structural complement to behavioral alignment approaches. While value-based alignment defines the normative objectives guiding AI systems, meta-alignment focuses on maintaining the structural conditions that allow systems to remain corrigible and responsive to corrective signals over time. The framework also introduces the concept of regime invisibility, where systems may appear aligned at the behavioral level while undergoing structural rigidification internally. This work is part of a broader research program including: Adaptive Closure in Agentic Systems (2026) Structural Observability and Governable Agentic AI (2026) Governing Governance: Structural Principles for Governing AI Acceleration Under Systemic Risk (2025)