Abstract

Large language models (LLMs) have become deeply integrated into human cognitive practice, yet the increasing variability of the AI infrastructure—characterized by frequent model updates, platform deprecations, and silent behavioral drift—produces a structured profile of cognitive disruption that existing frameworks fail to capture. This paper argues that classic extended mind theory and generic learning-cost accounts underfit this phenomenon by treating tools as static artifacts. To address this gap, we introduce the Interaction Schema S(H, T), a theoretical unit decomposing human–tool coupling into Protocol, Calibration, and State components. We develop a migration cost analysis organized around portability, drift, and repairability, demonstrating that disruption is not uniform but selectively targets specific coupling components. The framework generates discriminating predictions—including differential portability profiles across components and silent miscalibration under update drift—that are consistent with available behavioral evidence on cognitive offloading and tool transition costs. Our analysis extends extended mind theory from a static integration account to a dynamic reconfiguration account, introducing diachronic coupling as a necessary complement to synchronic integration conditions. Finally, we derive normative implications regarding the structural asymmetry between platform control over tool continuity and agents' cognitive dependence on that continuity, identifying portable components that agents can cultivate and strategies to mitigate disruption under conditions of uncontrolled tool change. This work provides the conceptual vocabulary and classificatory structure needed to investigate the continuity of cognition under variable cognitive infrastructure.