The Biggest AI Divide in Education Starts After Access

Generative AI is being sold to education systems as a way to democratize tutoring, feedback and personalized learning at scale. However, the same technology that can lift weaker performers may also reinforce the very inequalities it appears to reduce, especially when access, digital skills, language support and teacher capacity remain uneven.

A new paper "Does Generative AI Narrow or Widen Learning Gaps? The Divide Cascade: A Conceptual Framework for Equity, Access, and Quality Under Sustainable Development Goal 4," by Hasan M. Jamil of the University of Idaho, published in Sustainability, argues that GenAI is neither inherently equalizing nor inherently stratifying; its distributional impact depends on whether learners can move successfully from access to effective use, real learning gains and durable capability.

The global debate on AI in education has often treated access as the key equity problem. Jamil argues that this is too narrow. A student may technically have access to a chatbot and still derive less educational value because of weaker AI literacy, lower-quality language support, limited teacher mediation or greater dependence on AI-generated answers.

AI's Equalizing Power Is Real, but It Comes With Conditions

The case for GenAI as an equalizer is not simply speculative. The paper draws on workplace research showing that AI assistance can disproportionately benefit lower-skilled users, including one customer-support study where productivity rose by about 14% on average but by roughly 34% among the least-skilled workers. The implication is intuitive: AI can provide novices with knowledge, structure and support that more experienced users already possess.

Education offers a more directly relevant example. The paper highlights a six-week after-school intervention in Benin City, Nigeria, where students used GPT-4 in a structured, teacher-supported setting. The intervention produced a 0.31 standard deviation improvement on a composite assessment, while the cited cost-effectiveness analysis equated the gains to roughly 1.5 to 2 years of conventional schooling.

However, those gains did not emerge from unrestricted chatbot use. The Nigerian intervention was teacher-guided, curriculum-aligned and sustained over several weeks, while older evidence on intelligent tutoring systems similarly comes from carefully designed instructional environments rather than open-ended AI use. The study therefore cautions against assuming that access to a general-purpose model will reproduce the same outcomes.

If governments interpret strong results from structured interventions as proof that simply distributing AI tools will narrow learning gaps, they risk confusing the effect of the technology with the effect of the surrounding educational system. The study argues that the conditions around AI use may be as important as the model itself.

The Education Divide Does Not End When Access Is Equalized

The author proposes "divide cascade," a four-stage model that reframes educational inequality as a sequence of filters. First comes access to a capable AI system. Then comes effective use, followed by genuine benefit realization and, finally, durable learning that persists when the tool is no longer present.

Each stage can favor learners who already start with greater advantage. Wealthier students may have access to more capable subscription models, digitally fluent learners may know how to prompt and verify more effectively, high-resource languages may receive better model outputs, and well-supported schools may have teachers capable of integrating AI into instruction.

GenAI may compress performance differences among students who successfully clear every filter, while the filtering process itself excludes or weakens the gains of students with fewer resources. In other words, AI can look equalizing within a selected group of successful users while producing more stratified outcomes across the education system as a whole.

This is why access-only policies are unlikely to be enough. The framework predicts that providing devices, connectivity or free AI access may reduce the first layer of inequality while leaving later divides intact. AI literacy, teacher capacity, language quality and self-regulation can continue to determine who converts access into meaningful learning.

The paper also highlights a deeper problem: one intervention can narrow one type of inequality while widening another. In the Nigeria example, access and teacher support helped reduce a gender gap, while higher-baseline and higher-socioeconomic-status students still extracted greater gains, illustrating how different dimensions of advantage can move in opposite directions at the same time.

Better Answers Are Not the Same as Better Learning

Much of the strongest evidence for AI "leveling up" weaker users comes from workplace studies that measure finished outputs such as documents, code or customer-service tasks rather than whether users retain the underlying capability after AI assistance is removed. A student who produces a stronger essay with AI may appear to have improved, but that does not establish whether the student has developed better reasoning, writing or subject knowledge. The paper cites emerging research suggesting that AI can improve short-term task performance while encouraging cognitive offloading and reduced self-regulation.

The divide cascade introduces durability as the final and perhaps hardest test of equity. It predicts that gains measured immediately after AI-assisted work may weaken when students are assessed later and without assistance. If that effect is stronger among learners with less self-regulatory support, the apparent equalizing effect of AI could erode over time.

This is a major challenge for education systems racing to adopt AI. Schools and governments can easily measure access, usage and immediate performance, but those indicators may overstate success if they fail to capture retention, transfer and independent capability. For SDG 4, quality cannot be reduced to whether students produce better outputs while connected to a model.

Equity Will Depend on How Completely the System Is Engineered

The paper calls for capable models to be made available to disadvantaged learners, AI literacy to be explicitly taught, systems to be localized for different linguistic and cultural contexts, teachers to be trained, and pedagogy to preserve productive intellectual effort rather than automate it away.

Teacher capacity is particularly important. Where schools are well resourced, AI can augment teachers who verify, contextualize and structure its use. In weaker systems, however, AI risks being used as a substitute for scarce human instruction, leaving disadvantaged learners with automated assistance but less of the mediation that helps convert information into learning.

For developing countries, the stakes are especially high. The study notes that the existing evidence base is disproportionately concentrated in the Global North, higher education and English-language settings, while K-12 systems, low-resource contexts and low-resource languages remain underrepresented. That means the populations most central to SDG 4 are also those for whom the consequences of AI remain least certain.

The research also carries an important limitation: this is a conceptual framework, not a validated empirical model. The literature was selected purposively rather than systematically, the framework has not yet been tested with data collected specifically to evaluate its predictions, and its formal representation simplifies multidimensional disadvantage. These caveats mean the model should be treated as a testable agenda rather than a settled account of how AI will reshape education.

Its value lies precisely in sharpening what future research should measure. Jamil calls for studies that separate access, effective use, benefit realization and durable learning; track heterogeneous effects across different forms of disadvantage; test public-provisioning models; and assess whether AI narrows gaps on delayed, unaided learning rather than only on immediate performance.

For education policymakers, the broader lesson is that generative AI cannot be judged by adoption rates or average learning gains alone. A system can raise average performance while still widening the distance between learners who can fully exploit AI and those who cannot.