Generative artificial intelligence (GenAI) is frequently promoted as a democratizing force in higher education, capable of extending personalized instructional support to students who otherwise lack access to it. This review examines whether the empirical literature (2023–2026) supports that claim or whether it instead points toward a widening gap between students who benefit from GenAI and those who do not. Drawing on recent meta-analyses of learning outcomes alongside digital-equity scholarship, the review argues that GenAI's documented benefits for academic performance are conditioned on forms of access that are themselves unevenly distributed: reliable connectivity and devices, AI and digital literacy, and instructor capacity to design effective GenAI-integrated tasks. It organizes these conditions into a three-level framework of access, skill, and outcome divides, and reviews evidence on how each operates in higher education contexts. The review concludes that without deliberate institutional and policy intervention, GenAI risks reinforcing rather than narrowing existing educational inequities, and outlines equity-oriented implications for pedagogy, policy, and future research.