The regulation paradox: agentic AI, bounded autonomy, and self-regulated learning in higher education

Abstract

Agentic AI is moving educational technology beyond reactive assistance toward systems that can plan, remember, adapt, and initiate actions across time. In higher education, this development raises a basic pedagogical question: does greater AI autonomy help students become more capable self-regulated learners, or does it perform the regulatory work that students need to practice themselves? This Perspective uses self-regulated learning (SRL) to examine this question. It argues that agentic AI becomes educationally meaningful when its autonomy supports learning processes rather than only producing academic outputs. It also separates four risks within the regulation paradox: regulatory displacement, metacognitive laziness, reduced productive struggle, and epistemic over-trust. To address these risks, the paper proposes bounded autonomy as a pedagogical governance principle that limits the scope, timing, transparency, contestability, and duration of AI intervention. Bounded autonomy extends scaffolding and hybrid human-AI regulation by making decision authority explicit and by requiring regulatory responsibility to be returned gradually to learners. The paper further argues that responsible implementation depends on learner capability, teacher agency, institutional readiness, and evidence that regulatory skills transfer to tasks completed without AI. Agentic AI earns its educational value not by becoming more autonomous from learners, but by helping learners become more capable of learning without it.