The explicit teaching of critical thinking (CT) provides modest, uneven gains with limited transfer beyond the contexts in which it is taught. The authors argue that part of this difficulty is understanding how the different components involved in CT interact. The dominant frameworks describe CT as a set of cognitive skills, sometimes with dispositions added, without specifying how the components depend on one another. We propose a different framework, a layered model in which CT is an emergent, whole-system capacity: 1) intrapersonal foundations (for example emotion regulation and tolerance for ambiguity), 2) epistemic competencies (for example intellectual humility and open-mindedness), 3) cognitive subskills (for example analysis and inference), and 4) a composite layer at which CT is expressed and observable. The model predicts that weakness in any of the supporting layers will make critical reasoning fragile when stakes are emotional, social or identity-relevant, even when the subskills are intact. This offers a testable explanation for poor transfer and for the gap between demonstrated and authentic critical reasoning. Generative AI can now perform many of the operations that CT instruction and assessment emphasize, making the issue more pressing. Offloading may reduce the practice that builds these layers, and may also produce what we term confidence-by-proxy: mistaking fluent GenAI output for one’s own understanding. The authors outline implications for instruction, assessment and GenAI literacy, and then set out a research agenda. Research should begin with testing layer-specific interventions against skills-only instruction.