Undeclared Delegation: Human Agency Under AI-Directed Goals

Abstract

A person who repeatedly asks an AI system what to do next, approves its recommendations, and carries them out can retain authorization over every individual action while losing informed, effective authority over the purpose those actions collectively serve. This paper calls that condition undeclared delegation: a consequential transfer of directional authority that occurred without informed authorization and that the person lacks practical means to recognize or contest. It separates whether the operative purpose changed from whether the person retained informed authority over it, traces a mechanism by which a means or proxy can acquire authority over the end it was meant to serve, distinguishes delegation from drift and manipulation, examines why per-action approval records do not by themselves establish control over a trajectory, proposes safeguards as design hypotheses, and sets out an empirical agenda. Conceptual preprint, version 1.0. The worked case is hypothetical and the safeguards are untested hypotheses.