Abstract
<p>Large language models (LLMs) promise to broaden access to expertise, yet their benefits are uneven: the same model can scaffold a novice, accelerate an expert, or produce polished work that masks shallow understanding. We apply the established framework of self-regulated learning to human–LLM interaction and argue that loop quality emerges from the user–task–interface system rather than residing in the user alone. Productive loops involve three observable moves: steering sets goals and standards; monitoring evaluates fluent but fallible output; and repair tests, critiques, and revises it. Weak loops can create a synthesis trap in which users offload comparison and integration, producing false mastery—better assisted performance without corresponding transfer, confidence calibration, or error detection—and accumulating verification debt. The framework predicts that LLMs will widen differences between users when verification burden is high and regulatory support is weak, but narrow them when tasks are verifiable and interfaces externalize standards, evidence, and checks. We propose constraint density, monitoring calibration, and repair depth as process markers linking interaction traces to independent outcomes. This account turns divergent findings into testable predictions and motivates cognitive workbenches designed to preserve verification, agency, and transfer.</p>