Contemporary AI accountability discourse focuses on alignment, safety, bias, interpretability, and liability. This note introduces a missing governance category: deniable delegation. Deniable delegation is not an accidental byproduct of AI deployment but a structured opacity mechanism that allows a human principal to exert power, avoid attribution, escape regulatory visibility, and evade political accountability while publicly attributing agency to an AI system. This note defines the category, distinguishes it from existing frameworks, traces its primitive form in the 2023 Mata v. Avianca sanctions case, describes its sophisticated and stealth variants, and argues that the locus of governance must shift from the AI system to the human principal. It further identifies three architectural primitives—unilateral regulatory access, immutable audit trails, and freeze invariants—that function as anti-deniability infrastructure, converting deniable delegation into traceable delegation and collapsing the strategy.
Narnaiezzsshaa Truong (Sun,) studied this question.