Cognitive Load Theory (CLT) has been the dominant framework for instructional design research for over three decades, built on the principle that reducing extraneous load frees working memory capacity for the germane processing that produces durable learning. This principle was developed for static instructional environments in which the learner is the sole cognitive agent. The emergence of generative artificial intelligence (GenAI) tools introduces a dynamic information-processing partner whose contributions to the cognitive system are not adequately captured by CLT's tripartite architecture. Empirical anomalies are now accumulating in the GenAI-in-education literature — most sharply the Effort–Load Paradox, in which ChatGPT improves learning outcomes without altering perceived effort or cognitive load, and without germane load functioning as the operative mediating mechanism (Ramachandran, 2026). Drawing on a quasi-experimental study (N = 184) conducted in Indian management education and a synthesis of CLT literature in technology-enhanced learning contexts, this paper proposes two theoretical extensions: (1) dependency-induced extraneous load — the novel cognitive burden generated when AI scaffolding is abruptly withdrawn, manifesting most acutely at Bloom's application level; and (2) the efficiency pathway, through which GenAI improves outcomes by enhancing information quality rather than reducing load quantity, rendering its benefits invisible to standard rating instruments. These constructs are integrated with Vygotsky's Zone of Proximal Development, the Expertise Reversal Effect, and the Misinterpreted Effort Hypothesis into a revised theoretical framework for CLT in GenAI-mediated environments. Implications for instructional design, AI tool architecture, and CLT measurement methodology are discussed.