Sovereignty-Preserving AI Systems and Mechanisms: A Survey

This record contains a working-paper version of “Sovereignty-Preserving AI Systems and Mechanisms: A Survey.” The paper surveys technical and institutional mechanisms for preserving meaningful human and institutional control in AI-mediated environments. It organizes the literature across five layers of dependence: data, learning, action, exit, and ecosystem capacity. Mechanisms reviewed include on-device inference, federated adaptation, bounded agent architectures, machine unlearning, auditability, substitutability, and public compute. Using shared dimensions such as locality, participation in improvement, boundedness, reversibility, substitutability, and public verifiability, the survey argues that contemporary AI redistributes control across multiple technical boundaries rather than along a single axis such as safety or privacy. The resulting framework is intended to support the design and evaluation of governable AI systems.