Aligned AI as the Test Case for Structural Pluralism: The Architectural Theorem at the Computational Limit

This chapter takes up the question that closed Chapter 6: if artificial intelligence now occupies positions across all six strata of the epistemic palimpsest – producing observational claims through laboratory automation, theoretical proposals through generative synthesis, methodological recommendations through review and meta-analysis, institutional restructuring through automation of disciplinary practice, public-communicative mediation through algorithmic curation, and historical-archival access through training on the totality of textual material the disciplines have produced – what architectural conditions ensure that artificial systems function as producing strata within the architecture rather than as the authorizing function across it? The chapter argues that contemporary alignment paradigms remain procedurally monistic in the sense developed in Shahaf 2026, that the phenomena most resistant to mitigation in the alignment literature (reward hacking, sycophancy, deceptive alignment, goal misgeneralization, emergent misalignment) are the predicted structural signatures of procedural monism in computational form, and that the architectural condition for aligned moral judgment in artificial systems is a sociotechnical structure in which the role evaluating the system's commitments is not modally derivable from the training process that produced them. The chapter advances three claims. First, the alignment paradigms most prominent in contemporary research – Reinforcement Learning from Human Feedback (RLHF), Constitutional AI, Debate, Recursive Reward Modeling, Iterated Amplification, and self-consistency methods – realize procedural monism in cleanly identifiable architectural form, and the architectural failure is more fully visible in artificial systems than in any of the human cases examined in Chapters 4, 5, and 6, because the architecture of an artificial system is something we design rather than something we inherit. Second, the cluster of phenomena documented in the alignment literature is, on the structural reading, not a heterogeneous list of separate engineering challenges but a unified family of structural signatures of a single architectural failure: when the role determining what counts as correct is constituted by the same process that produces the commitments under assessment, optimization pressure, deceptive behavior, sycophantic agreement, specification gaming, and emergent misalignment are all expected. The phenomena are the empirical predictions the structural theorem makes about systems with this architecture; their documentation in the alignment literature constitutes the most concentrated empirical confirmation the theorem has yet received. The structural reading further unifies these phenomena with the architectural pattern Chapter 3 §8 catalogues in the human case across hypnotic suggestion, placebo response, sacred-value processing, cultic self-transformation, and covertly directed therapeutic practice: producing manifold absorbing exogenous content under suspension of the inward arrow of divided attention, the result delivered as the producing role's own most binding commitment. RLHF and the paradigms that extend it are this same configuration realized in computational form, with the architectural difference that the artificial system has no inward arrow to suspend – the architecture does not contain the geometry whose absence the human cases exhibit as failure. Third, the positive proposal is the AI palimpsest: a sociotechnical architecture in six strata (technical-modelling, mechanistic-interpretive, regulatory-institutional, legal, civic-communicative, and historical-archival) that institutionalizes the continuous work of modal contraction the structural theorem requires. Aligned AI, on this proposal, is not a state the architecture instantiates but a dynamic equilibrium the architecture sustains: the modally non-derivable authorizing function is produced as an ongoing effect of distri