Why Aligned AI Requires Structural Pluralism

This paper extends the structural argument of No Self-Authorizing Norms (Shahaf, 2026a) to the problem of AI alignment. I argue that alignment cannot be secured merely by specifying the right values, preferences, or constitutional principles, because moral judgment requires structural plurality: an evaluative authority whose standpoint is not modally fixed by the commitments it assesses. The claim concerns systems entrusted with framework-level moral verdicts in open normative settings, not every bounded form of safe or compliant behavior. The specified implementations of current alignment paradigms – RLHF, Constitutional AI, Debate, Recursive Reward Modeling, and self-consistency methods, in their deployed or AI-mediated, frozen-judge forms – are procedurally monistic wherever every authority role bearing on a verdict is either a fixed function of the commitments under assessment or constant across the admissible authority-configurations; there, the practice is diagnostically fixed for those verdicts, however functionally distinct its commitment-generating and evaluative mechanisms remain. This structure helps explain reward hacking, sycophancy, deceptive alignment, goal misgeneralization, and specification gaming as related expressions of the same architectural deficit, and reads emergent misalignment as the higher-order form in which they converge. The paper presents modal non-derivability as a necessary, though not sufficient, condition for aligned moral judgment, and argues that genuine AI alignment must be assessed at the level of the sociotechnical practice rather than as a property of a model alone, even where the requisite plurality of roles is internalized within a single system. The condition is then translated into a multi-rate research program that separates a construction-side test of the authority relation from an empirical test of operational capacity. A developmental account locates the same condition in the formation of the human subject, and a corollary concerning the observer shows that whether a given output of a condition-satisfying system is the exercise of framework-level judgment or its imitation cannot, in general, be inferred from behavior alone.