The Explanation Fallacy: Why "Faithful Explanations" Cannot Serve as a Governance Primitive for AI Systems

Current AI governance discourse increasingly demands “faithful explanations” from large language models (LLMs). This requirement presumes that (1) human explanations are faithful, (2) natural language is a reliable container for cognition, and (3) post-hoc narratives can serve as admissible evidence of decision formation. None of these assumptions hold. Human cognition is reconstructive, not transparent; language is lossy and metaphorical; and LLMs inherit these properties from their training data. This note argues that explanation-centric governance is structurally incapable of producing admissible decision authority. Governance must shift from post-hoc narrative extraction to substrate-level constraints that shape intent formation, privilege activation, and admissibility before a decision exists — making governance constitutive rather than interpretive.