UNESCO Report: AI Tutoring Closes Learning Gaps in 12 Developing Nations

Published 2025-06-15 · Education AI

A landmark report released by UNESCO on June 14, 2025, presents the most comprehensive evidence to date that AI-powered adaptive tutoring can meaningfully close learning gaps in low- and middle-income countries. The two-year study, conducted across 12 nations in sub-Saharan Africa, South Asia, and Southeast Asia, found that students using AI tutoring platforms achieved literacy gains 34% larger than peers receiving standard classroom instruction alone.

The report, titled "Artificial Intelligence and the Global Learning Crisis: Evidence from 4,200 Classrooms," was produced by UNESCO's Global Education Monitoring team in collaboration with researchers from the University of Cambridge, the Indian Institute of Technology Bombay, and the African Institute for Mathematical Sciences. It draws on data from 4,217 classrooms encompassing approximately 312,000 students aged 8 to 14.

Study Design and Methodology

The research employed a stepped-wedge cluster randomised design, considered the gold standard for evaluating educational interventions at scale. Schools were randomly assigned to receive AI tutoring in one of three sequential phases, allowing researchers to compare outcomes within the same schools before and after the intervention while controlling for temporal trends.

The 12 participating countries were selected to represent a range of educational contexts: India, Bangladesh, Nepal, Kenya, Tanzania, Uganda, Ghana, Nigeria, Senegal, Rwanda, Cambodia, and Laos. Within each country, between 30 and 50 schools were enrolled, stratified by urban-rural location and baseline student performance.

Four AI tutoring platforms were evaluated: Mindspark (India), Eneza Education (Kenya and Tanzania), a localised version of Khan Academy's Khanmigo deployed in Cambodia and Laos, and a UNESCO-developed open-source platform called PALF (Personalised Adaptive Learning Framework) used in the remaining countries. All four platforms shared core features: diagnostic assessment, adaptive content sequencing, real-time feedback, and progress dashboards accessible to teachers.

Key Findings

The headline result — a 34% improvement in literacy gains relative to the control condition — was measured using the Early Grade Reading Assessment (EGRA), a standardised oral evaluation of letter knowledge, word reading, passage reading, and reading comprehension. Students in the AI tutoring group improved their EGRA scores by an average of 18.3 scale points over the two-year study period, compared with 13.7 points in the standard instruction group.

Numeracy results were even more striking. On the Early Grade Mathematics Assessment (EGMA), AI-tutored students gained 22.1 scale points versus 14.8 in the control group — a 49% relative improvement. The gap was largest for foundational arithmetic operations (addition, subtraction, multiplication) where adaptive platforms excelled at identifying and remediating specific misconceptions.

Importantly, the largest gains were concentrated among students who started furthest behind. Children in the bottom quartile of baseline performance improved by 41% more with AI tutoring than their peers in the control group, compared with a 19% improvement for students in the top quartile. This inverts the common concern that technology-based interventions primarily benefit already-advantaged learners.

Teacher Engagement and Classroom Integration

A persistent question in AI education research is whether adaptive software complements or undermines teacher-led instruction. The UNESCO study found that the most effective implementations were those where the AI platform was used as a supplement — typically 30 to 45 minutes per day, three to four days per week — while teachers used the platform's analytics dashboard to inform their whole-class instruction.

In schools where teachers received structured training on interpreting platform data and incorporating it into lesson planning, student gains were 28% larger than in schools where the platform was simply distributed without teacher orientation. This finding aligns with broader research on blended learning, which consistently shows that technology amplifies good teaching but does not compensate for its absence.

Dr. Vongai Nyika, the report's lead author, emphasised this point: "The platforms did not replace teachers. The best results came from classrooms where teachers used AI-generated insights to target small-group interventions, differentiate homework assignments, and identify students who were silently disengaging. The technology made visible what teachers suspected but could not always quantify."

Cost and Infrastructure Considerations

One of the report's most practically relevant findings concerns cost-effectiveness. The average per-student cost of deploying AI tutoring across the 12 countries was $8.40 per year, inclusive of hardware, connectivity, platform licensing, teacher training, and maintenance. This compares with an estimated $340 per student per year for the average low-income country's total education expenditure. In other words, the intervention added roughly 2.5% to baseline education spending.

Infrastructure varied significantly across sites. In urban India and Kenya, most schools had reliable electricity and intermittent internet connectivity. In rural Rwanda, Laos, and northern Nigeria, the platforms ran on offline-capable tablets preloaded with content, synchronising with cloud servers when connectivity was available — typically once per week via satellite or mobile data links provided by the project.

The offline-first deployment model, built into the PALF platform, proved critical to the study's success in remote areas. Content updates, student progress data, and adaptive model refinements were packaged as compressed payloads that could be transferred over extremely low-bandwidth connections. Dr. Joaquin Navarro, who led the PALF engineering team, noted that the system was designed to function for up to 30 days without any internet connection.

Equity and Inclusion Challenges

The report does not shy away from limitations. Gender disparities persisted in several countries: in northern Nigeria and parts of rural Nepal, girls' access to shared devices was constrained by cultural norms around technology use, and the AI platforms — which did not initially account for gendered patterns of device access — inadvertently reinforced these gaps in early implementation phases. Corrections were made mid-study by introducing female-only device sessions and recruiting female learning facilitators.

Language coverage was another constraint. While the platforms operated in major national languages (Hindi, Swahili, Khmer, French, English), students whose first language was a local or indigenous language often struggled with the instructional content. UNESCO recommends that future deployments invest in multilingual content creation and leverage recent advances in speech-to-text and machine translation to support mother-tongue instruction.

Disability access was largely unaddressed. Few platforms included screen reader compatibility, sign language support, or content adapted for learners with cognitive disabilities. The report calls this a "significant ethical gap" and urges developers to adopt universal design principles from the outset rather than retrofitting accessibility features after deployment.

Policy Recommendations

UNESCO's policy recommendations are directed at governments, development agencies, and platform developers alike. The report urges ministries of education to develop national AI-in-education strategies that include clear standards for data privacy, algorithmic transparency, and teacher professional development. It calls on development funders to support open-source, interoperable platforms rather than proprietary systems that create vendor lock-in. And it challenges platform developers to invest in multilingual, offline-capable, and disability-inclusive designs.

The findings complement those of a separate UNESCO analysis of Khan Academy's Khanmigo platform, which recently reached 10 million students worldwide. Together, the two reports build a compelling case that AI tutoring, when deployed thoughtfully and equitably, can address a learning crisis that affects an estimated 617 million children and adolescents globally who lack minimum proficiency in reading and mathematics.

Explore our AI in Education research repository for broader coverage of adaptive learning, intelligent tutoring systems, and education policy. For related work on AI-driven health interventions in low-resource settings, see WHO's global guidelines on AI in health.

← Back to Publications