Cognitive Sovereignty in the Age of Educational AI: A Second-Order Cybernetic Framework for Pedagogical Auditing, Human Agency, and Public Infrastructure

Educational artificial intelligence should be evaluated not only by the quality of the work it helps students produce, but also by the capabilities students retain, develop, and exercise when its assistance changes or disappears. This white paper develops a second-order cybernetic framework for distinguishing task completion from durable learning and for translating that distinction into public oversight. It treats the learner, teacher, AI system, assessment regime, and procuring institution as an interconnected system whose incentives and measurement practices shape one another. Selected experimental evidence indicates both risks from unrestricted answer provision and benefits from deliberately structured tutoring; neither finding warrants a universal claim that AI inevitably diminishes or enhances cognition. The paper proposes a Pedagogical AI Auditing Agency (PAIAA), understood first as a public auditing function and, where justified and lawfully authorized, as a dedicated institution. Its architecture combines deployment-specific assurance cases, adaptive scaffolding, delayed retention and transfer assessments, assisted verification tests, privacy-preserving evaluation, versioned certification, and independent scrutiny of the auditor itself. A formal conceptual model, a conditional coordination-game analysis, falsifiable hypotheses, and a staged research protocol make the proposal open to empirical revision. Public-option infrastructure is advanced as a means of preserving institutional exit and contestability, not as a guarantee of epistemic neutrality. Cognitive sovereignty is defined as the distributed capacity to understand, challenge, redirect, and replace the systems through which knowledge is produced. The central policy claim is normative but testable in implementation: public education should purchase demonstrable capability development rather than mistake fluent output, engagement, or adoption speed for learning. Author’s Deposit & Verification Note (v1.0): Pursuant to the pre-deposit declarations in Appendix E.2 and E.3, the named author (Dustin Cleanwell Liu) confirms completion of pre-deposit review for this Version 1.0 working paper, assumes authorship responsibility for its public deposit, and declares no competing financial interests. In Section 6.1 (page 13), minor inline glyph rendering artifacts in the PDF text refer to the instructional context ({c_t}), the observable work product ({z_t}), and the instructional policy ({π}), matching Equations (1)–(4).