AI Governance and the Architecture of Dependency

A system does not have to remove choice in order to shape it. It can define the problem, structure the options that appear, and create conditions under which a voluntary choice becomes costly to reverse. These mechanisms predate artificial intelligence. This paper asks what changes when AI-mediated systems operate them at greater speed, scale, and adaptability, and when the people subject to them cannot practically inspect, contest, or leave.

The paper develops a four-part lens (problem definition, choice architecture, dependency formation, reversibility) and distinguishes institutional dependency from epistemic dependency. It adds an agency and accountability audit of six questions: who sets the objective, who controls the data, model and infrastructure, who can inspect the evidence, who bears the cost of error, who can override or replace the system, and what alternatives remain. Its central hypothesis, an accountability asymmetry, is stated with the evidence that would weaken it.

A rubric scores safeguards by whether they have been exercised, not merely documented, and is applied to three current arrangements: the Robodebt scheme, the Medicare WISeR model, and the UnitedHealth nH Predict litigation. Where public evidence allows no score, the gap is reported as a finding. Appendices provide a one-page citizen audit card and a legal and case status snapshot as of 21 October 2026.

This is a working paper. It has not been peer reviewed and is not legal advice. Claims are labelled documented, reported, interpretation, or hypothesis. The rubric is unvalidated and was scored by a single rater from public sources. Companion to Johnson (2026a) and Johnson (2026b).