The Global AI Classroom Gap

The UN’s Sustainable Development Goal 4, quality education for all by 2030, was already clearly unachievable before AI. The World Bank’s 2022 learning poverty report found that 70% of ten-year-olds in low- and middle-income countries could not read and understand a simple text. In sub-Saharan Africa, that figure was 89%. These are not countries with access problems, in the sense that most children in those countries are in school. They have quality problems: school attendance is high, and learning is low, because teachers are undertrained and overburdened, materials are inadequate, and the educational infrastructure doesn’t deliver instruction that actually produces reading ability.

AI optimists point to this situation and see opportunity. If AI can deliver high-quality instructional content via a tablet to a student in a Lagos classroom whose teacher has a third-grade education and 60 other students to manage, surely that’s a win? Surely that’s democratization?

The argument is not wrong. But it’s incomplete in ways that matter for how resources should be allocated.

The first problem is infrastructure. The students who are furthest behind educationally are disproportionately in contexts with unreliable electricity, low internet bandwidth, expensive data, and limited device availability. These are solvable problems, and there are genuine efforts, Google’s Project Loon, Starlink’s education access programs, the $30 OLPC tablet effort, aimed at solving them. But “solvable” and “solved” are different things. As of 2025, approximately 2.6 billion people remain without internet access, disproportionately in the low-income countries where educational deficits are worst. AI that requires consistent connectivity and a reasonable device is not an immediate intervention for this population.

This isn’t an argument against working on access. It’s an argument against treating access as the primary barrier when the primary barriers are the ones that would remain even with full access: teacher quality, educational culture, parental education levels, economic pressure that keeps children out of school, nutrition, and health. A child who can’t read because she’s suffering from untreated anemia, or because her parents need her to work, or because her classroom has 90 students and one teacher, this child’s situation is not solved by an AI app on a tablet, even a free one with offline capability.

The second problem is more structural and more important to get right. When AI improves educational outcomes, it improves them proportionally. That is, the absolute gains tend to be larger for students who are already above a threshold of basic literacy and motivation. This is not a critique of AI specifically; it’s a property of most educational interventions. The educational research term is “Matthew effects,” after the Biblical principle that to him who has, more shall be given. Students with stronger foundations absorb new educational inputs better because they have more schema to attach new learning to.

AI tutoring systems that work by adapting difficulty and providing explanation require a student who can read the explanation. A student who cannot read at all is not helped by a better-worded explanation. A student who cannot do basic arithmetic is not helped by an AI that provides worked examples of algebraic problems. The baseline competencies that make higher-order educational inputs effective, literacy, basic numeracy, the habit of engaging with written material, are exactly the competencies that are missing in the populations with the worst educational outcomes.

The gap between “baseline for AI effectiveness” and “current state of the worst-off students” is large. Bridging it is a human-intensive, low-technology intervention: trained teachers, in classrooms with manageable ratios, doing the foundational work of teaching children to decode text and understand number. AI can augment this work; it cannot substitute for it at the foundation.

What does this mean for where AI investment in global education actually helps? It’s more specific than the broad “AI for all students” framing suggests.

AI has the most immediate impact in the middle of the global distribution: students in middle-income countries who are literate and have basic numeracy but lack access to high-quality instruction in secondary school subjects. These students, in countries like India, Brazil, Indonesia, Mexico, Nigeria’s urban population, are above the threshold where AI can help and below the threshold where human tutoring is economically accessible. Byju’s early growth in India, before its financial crisis, documented this market clearly: tens of millions of students in Indian secondary schools who had the baseline skills to use a tablet-based learning product and who were motivated to improve their results on high-stakes examinations. AI tools for this population are genuinely powerful.

India’s National Education Policy 2020 explicitly acknowledged AI’s potential for this segment, and the EdTech ecosystem that grew around JEE and NEET exam preparation, the entrance exams for Indian engineering and medical colleges, demonstrated that technology-assisted learning at scale was possible. The problem wasn’t that the technology didn’t work. It was that the business model of for-profit EdTech didn’t serve students without means, and the government platforms that could serve them weren’t built quickly enough or with enough quality.

There’s a language problem underneath the infrastructure problem that rarely gets discussed. Most AI educational tools of high quality are in English. The large language models that power AI tutoring systems were trained disproportionately on English text, and their quality drops significantly in lower-resource languages. A student in rural Tamil Nadu studying in Tamil, a student in Ethiopia studying in Amharic, a student in Vietnam studying in Vietnamese, the AI educational tools available to these students are substantially worse than those available to students studying in English, even controlling for device access.

The investment required to close this gap is not primarily a problem of distribution. You can’t fix it by giving everyone a tablet. It requires building high-quality AI training pipelines for languages that currently lack them, which requires data collection, annotation, and model development in those languages. This is expensive and produces tools that serve large numbers of people but generate limited commercial revenue in markets with low purchasing power. It will not happen primarily through market mechanisms.

The version of AI in global education that is actually achievable in the next decade, if resources and priorities are aligned correctly, looks like this: AI that raises the effectiveness of teacher training, not student-facing AI that bypasses teachers. The bottleneck in educational quality in most low-income countries is teacher quality, not student access to content. A teacher training system augmented by AI, that can identify specific gaps in a teacher’s subject knowledge, provide targeted instruction on common student misconceptions, help teachers design better formative assessments, and provide feedback on classroom practice, would help far more students than the same investment in direct AI tutoring products.

This is a less exciting pitch than “AI tutor for every child.” It doesn’t produce a product you can put in a press release with a photo of a child on a tablet. It requires working with governments, teacher unions, and educational bureaucracies in politically complex environments. It’s slower. It’s also more likely to produce durable improvements in educational outcomes, because the classroom teacher is still the primary delivery mechanism of education for most of the world’s students, and will remain so for decades regardless of what AI can theoretically deliver.

The global AI classroom gap is real, and it is going to widen before it narrows, not primarily because of access, but because of the structural asymmetry in AI’s effectiveness relative to educational baselines. Acknowledging this clearly is the first step toward investing in the right interventions.

The funding question is central and gets systematically avoided in optimistic framings of AI’s potential in global education. The organizations with the resources to fund AI educational interventions at scale — the World Bank, USAID, the Gates Foundation, the Hewlett Foundation — have all made public commitments to AI in education. The combined annual disbursement of those commitments, as of 2025, is approximately $800 million globally. UNESCO estimated in 2023 that addressing the global learning poverty problem through conventional means (more trained teachers, better materials, infrastructure) would require approximately $97 billion annually above current spending. AI is not going to bridge a gap of that magnitude at $800 million.

The more realistic version of AI’s contribution to global education equity is additive at the margins: making modestly better resources available in contexts where somewhat better resources can make a real difference, while the structural interventions, political, economic, public health, that actually determine educational outcomes remain underfunded. This is valuable. It’s not the transformational democratization story that gets told in conference presentations.

One specific intervention that deserves more attention than it’s receiving: AI for early literacy assessment. Identifying which children are not learning to read, and which specific decoding or comprehension problems they’re encountering, is a prerequisite for targeted intervention. In most low-income countries, teacher capacity for systematic literacy assessment is limited; a single teacher with 50 students doesn’t have the bandwidth for individual oral reading assessments. Low-cost AI systems that can assess a child’s reading level through a short voice interaction, identify specific error patterns, and flag children for teacher attention are genuinely within current technical reach and could be deployed at a cost that development organizations can fund.

This kind of targeted, modest, evidence-based AI intervention is less exciting than “AI tutor for every child” but considerably more achievable and considerably more likely to produce real educational improvements in the contexts where they’re most needed. The global AI classroom gap will narrow when serious people allocate serious resources to the boring, hard, contextually specific work of improving educational quality in low-income countries. AI can help. It’s not a substitute for doing that work.

The test of whether the AI-in-global-education conversation is serious is whether the resources follow the evidence. Teacher training, early literacy assessment, foundational numeracy: these are the interventions with the strongest evidence base and the weakest funding relative to their potential impact. AI-powered tutoring platforms for middle-income students with smartphones: this has the best business model and the weakest evidence of impact on the students most in need. If the next five years of global EdTech investment continues to follow the money rather than the evidence, the AI classroom gap will widen. Not because AI is bad for education, but because the global education system will have used AI to serve the students who needed it least, one more time.