The Unceded Intelligence Test: Can You Limit, Challenge and Revoke Your AI?
Description
This paper develops the Unceded Intelligence Test as a proposed framework for assessing whether people retain meaningful authority when delegating tasks to artificial intelligence. It examines four domains: authorization, scope, contestability and revocation. Together, these address who grants an AI system its mandate, what limits its powers, whether consequential actions can be challenged and corrected, and what happens when permission is withdrawn.
Each domain receives a finding of demonstrated, incomplete or failed, supported by evidence and clearly stated assessment boundaries. The proposed decision rule requires all four domains to be demonstrated for an arrangement to pass. Failure in one domain cannot be offset by strength elsewhere, while missing evidence remains distinguishable from an established failure.
The paper outlines assessment preparation, evidence collection, controlled exercises, reporting and reassessment. A hypothetical purchasing-assistant example illustrates how successful task performance can coexist with failed revocation. An appendix provides a suggested structure for recording assessments.
This expanded repository edition connects the Unceded Intelligence framework with scholarship on meaningful human control, human–AI interaction, automation and algorithmic accountability. The paper is a conceptual and methodological proposal: it reports no empirical validation, and its findings should not be interpreted as comprehensive certification of AI safety or ethical acceptability.