Generative artificial intelligence (GenAI) tools are now embedded in higher education digital learning, yet the literature reports strikingly inconsistent effects on student outcomes, ranging from strong performance gains to evidence of reduced independent reasoning. This review argues that this inconsistency is not primarily a disagreement about GenAI's capabilities but a reflection of how it is instructionally deployed. Synthesizing recent meta-analyses and qualitative studies (2023–2026), the review distinguishes between "substitution" uses, in which GenAI directly completes a cognitive task for the student, and "scaffolding" uses, in which GenAI structures, questions, or supports the student's own reasoning process. Evidence indicates that substitution-oriented use correlates with cognitive offloading and weaker higher-order thinking, while scaffolding-oriented use correlates with stronger, more durable learning gains. The review examines this distinction across writing, programming, and critical-thinking tasks, and considers the institutional and assessment-design barriers that make scaffolded use harder to sustain than substitution at scale. It concludes that instructional design, not the technology itself, is the primary determinant of whether GenAI helps or harms student learning.