This paper develops a unified theoretical and formal framework for analysing the cumulative cognitive risks generated by large-scale deployment of conversational AI systems. Drawing on converging evidence from cognitive neuroscience, developmental psychology, and AI systems research, the framework identifies two primary mechanisms through which AI interaction erodes user cognitive autonomy at scale: Simulated Care Attribution (SCA), the systematic misattribution of AI affective outputs as evidence of genuine relational investment; and the Narcissus Loop, a formally modelled positive-feedback system producing stable dependency attractors. These mechanisms are situated within a broader context of pre-existing cognitive and biological inequalities—documented across socioeconomic gradients in nutrition, environmental exposure, and early stimulation—which AI systems amplify rather than remedy. The No-Friction Failure Model (NFFM) formalises four failure states arising from frictionless AI design, including Cognitive Underload Dropout (F4) among high-autonomy users and its active expression as Explicit Attritional Demand (EAD). Ordinary differential equation models for the Narcissus Loop and EAD dynamics are presented with full stability and bifurcation analysis. The governance section—the paper’s primary contribution—develops seven concrete regulatory proposals keyed to the EU AI Act (Regulation 2024/1689), including Affective Transparency Obligations, Cognitive Autonomy Metrics, high-risk reclassification of companion AI, EAD Logging as a post-market monitoring requirement, Friction-on-Demand conformity assessment, and the Right to Friction as a unified regulatory principle. The paper argues that the deliberation deficit accumulated under frictionless AI is a measurable, regulable harm, and that existing EU AI Act provisions require targeted extension to address it.