The Pocket Tutor Has a Price: Scaffolding Self-Paced Learning with AI

The dominant narrative casts large language models (LLMs) as a low-cost realization of Bloom’s two-sigma tutor: patient, always available, personalizing at scale. Self-paced online learning, the fastest-growing segment of higher and continuing education and the backbone of the data-science workforce pipeline, is where this promise is most attractive and most dangerous. Through an integrative synthesis of cognitive science, learning theory, computing-education evidence, and the economics of services, we argue that self-paced settings strip away the implicit scaffolding of the classroom (instructor presence, peers, accountability, ambient feedback), so unguided LLM access defaults to cognitive offloading, metacognitive disengagement, and fragmented learning that erodes durable competence. Recent large-scale evidence, including a survey of over ninety-five thousand undergraduates that found sharply elevated failure rates in introductory computing, is consistent with this. Rather than reject LLMs, we ask what a protective, holistic environment must do, grounding the answer in cognitive load theory, self-regulated learning, desirable difficulties, the zone of proximal development, and the ICAP framework, and we translate it into a layered reference architecture organized around a persistent learner model: longitudinal tracking, a curriculum coherence graph, adaptive personalization, cognitive gating, reflective agents, and process-based assessment with human checkpoints. Our central claim is economic: the configurations that protect cognition are far more resource-intensive—in compute, data and provenance infrastructure, authored artifacts, money, time, and human labor—than the near-free configuration that harms it. We characterize this cost asymmetry and its equity and workforce consequences and argue that resourcing the protective environment is, in effect, educational and workforce policy.