Human Approval, Delegation and Accountability in Consequential Systems

Abstract

EvidenceToEffect E2E-17 develops a semantic model for human approval, delegation, oversight, intervention, and accountability in consequential systems, including AI-agent contexts. The paper addresses a common analytical problem: “human in the loop” is often treated as a sufficient answer to autonomy risk even though the phrase may refer to several different claims. A human may have seen information, understood it, approved a proposal, delegated action, retained the ability to intervene, or remained accountable afterward. These claims are related but not equivalent. E2E-17 therefore distinguishes six human-governance claims: awareness, understanding, approval, delegation, intervention/oversight, and accountability. A human click, signature, review, or presence is evidence of some human involvement, but it does not automatically establish informed approval, current effect authority, effective intervention capability, or complete accountability for the realized outcome. The paper emphasizes that delegation does not erase effect specificity. A principal may delegate a purpose, role, budget, time window, or class of actions while the exact consequential effect remains subject to current scope, constraints, context, and authority at execution time. It also identifies recurring semantic failure patterns in human oversight, including rubber-stamp approval, scope drift, latency drift, interface ambiguity, authority ambiguity, and accountability diffusion. Worked examples cover agentic purchasing, software deployment, high-risk AI decision support, and physical automation. These examples illustrate why nominal human approval or override capability should not automatically be treated as evidence of effective oversight or continuing authority for a later consequential effect. The paper positions EvidenceToEffect alongside Article 14 of the EU AI Act, NIST AI RMF governance outcomes, and current NIST work on software and AI-agent identity and authorization, while preserving their distinct legal and governance scopes. This publication is intentionally implementation-agnostic. It defines no approval interface, delegation token, authorization protocol, override design, accountability score, legal-duty allocation, technical human-in-the-loop architecture, or conformity method. Series: EvidenceToEffect Research Series · E2E-17Version: 1.0.0Author: Ho Wa KUPublication date: 3 October 2026Foundational reference: EvidenceToEffect v1.0.0 — DOI: 10.5281/zenodo.23040907