Does Generative AI Narrow or Widen Learning Gaps? The Divide Cascade: A Conceptual Framework for Equity, Access, and Quality Under Sustainable Development Goal 4

Generative artificial intelligence (GenAI) is being absorbed into education as a new infrastructural layer, promising individualized tutoring, instant feedback, translation, and accessibility support at marginal cost. Sustainable Development Goal 4 (SDG 4) calls for inclusive and equitable quality education for all, yet the evidence on whether GenAI advances or undermines that goal points firmly in both directions at once. At the task level, GenAI and intelligent tutoring systems repeatedly compress performance distributions, with the largest gains accruing to lower-skilled and lower-baseline participants. At the system level, a parallel literature on access, AI literacy, language, disability, teacher capacity, and over-reliance finds that benefits are conditioned by resources that track prior advantage. These two literatures are usually read as being in tension, and the tension is usually resolved by privileging one of them. This article argues that both are correct and that the appearance of contradiction is an artifact of conflating distinct stages of a single pathway. We develop the divide cascade: a four-stage filter model—access, effective use, benefit realization, and durable learning—in which each stage has a pass rate that correlates with prior advantage. Because pass rates compound multiplicatively across stages while compression acts additively within a stage, a technology can compress outcomes among those who clear every filter and still stratify outcomes across the population as a whole. We formalize this structure, derive the condition under which the stratifying force dominates the equalizing one, and state three predictions that distinguish the cascade from an access-centred account: that access-only interventions should attenuate rather than close benefit gaps, that a single intervention can narrow one gap while widening another simultaneously, and that measured equity gains should decay as the evaluation horizon lengthens. We also specify what would falsify the model. The framework is then used to identify the conditions that set the sign of GenAI’s distributional effect, to map those conditions onto SDG 4 targets, and to derive a testable research and policy agenda. A recurring corollary is methodological: the strongest evidence for compression comes from workplace-productivity studies that measure produced artifacts rather than durable learning, so its transfer to education is an open question that the cascade locates precisely rather than assumes. GenAI, we conclude, is neither inherently an equalizer nor an amplifier; it is a multiplier whose sign is set by how completely the cascade is engineered for the learners who start behind.