Residual governance for responsible embodied AI: authority, contestability, and cognitive sovereignty

Abstract

Artificial intelligence (AI) ethics increasingly concerns systems that do not merely recommend but intervene: clinical decision support, human–machine teaming, robotics and automated handover protocols can shift timing, attention and authority under uncertainty. This article argues that responsible embodied AI requires residual governance: an explicit account of which approximation errors remain, which are safe to tolerate and who may act on them. The paper develops two linked concepts. A residual ledger records action-relevant remainders across physical, causal, human-state, latency, transfer and energy dimensions. An authority ledger records who converts residual evidence into intervention, on what basis, with what uncertainty and with what opportunity for contestation. This framework extends familiar themes in bounded rationality, active inference and safe reinforcement learning by treating authority, contestability and oversight cost as first-class ethical variables rather than external governance add-ons. It uses physical AI, physiological computing and human-AI handover as pressure cases, and proposes evaluation through action calibration, dual-task interference and auditable authority transitions. The central claim is that cognitive sovereignty is preserved not by eliminating all algorithmic influence, but by making influence accountable, contestable and reversible when AI systems act under approximation.