Tracing AI Action Back to Human Authority: Authority Provenance, Authorization, and Revocation in Distributed AI Execution
Year: 2026
Type: preprint
Abstract: Microsoft AI’s September 2026 draft Humanist AI Code of Conduct proposes meaningful human control through instruction hierarchy, bounded delegation, correction and shutdown. This conceptual and comparative paper asks what evidence is required to reconstruct whether a consequential action remained within the recognized institutional authority of the actor entitled to authorize it. Drawing on Institutional Authority and Governance Dynamics (IAGD 2.3), it presents Authority Provenance as a cross-cutting evidentiary assessment rather than an independent framework or technical enforcement invention. The procedure specifies six questions—Origin, Scope, Lineage, Standing, Correction and Propagation—and separates evidence differences, governance findings and authorization decisions. It places these questions in dialogue with public-administration accountability scholarship, AI auditing, technical provenance, access-decision standards and agent-delegation security research. Two complementary, nonmatched secondary-record applications examine the 2025 Replit production-data incident and the 2026 ROGUE experimental delegation benchmark. Source-to-finding identifiers and a strong integrated-governance rival constrain the conclusions. The analysis illustrates how to ask and bound institutional-authorization questions, but neither case provides a complete institutional mandate and execution trace. No uniquely superior diagnostic capability, empirical validation, independent reproduction, or Microsoft adoption is established. The contribution is a specified, falsifiable documentary assessment procedure and its limitations.
Author: Davis, Daron L.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212690164