Technology is often touted as a panacea for education in low- and middle-income countries. But it must be supported by the right government policies and sensitivity to local educational and cultural realities.
The advent of generative artificial intelligence (AI) has sparked a new wave of tech utopianism, with some arguing that the technology will fundamentally revolutionize teaching and learning. However, the success or failure of AI, and more broadly technology in education, might depend less on infrastructure and computing power than on policy and cultural relevance. The Global Education Evidence Advisory Panel noted that technology interventions that solely focus on infrastructure or hardware have not worked to improve learning outcomes.
As professor of education Larry Cuban sees it, education innovations are often oversold but under-contextualized. In many low- and middle-income countries, devices and software are often pitched as panaceas to sundry education issues, with insufficient consideration given to country or school norms, policies, stakeholders, or culture. When this happens, proposed solutions might falter or, worse, harm rather than improve teaching and learning.
This is not a call to rule out AI’s role in education outright. Today, over 80% of students in sub-Saharan Africa can’t read by age 10; there is a need for bold initiatives that leverage AI’s potential to automate routine tasks, support teachers’ professional development and provide personalized scaffolding to students. As we see below, World Bank Group-supported initiatives show how policies and culture must be baked into AI interventions in education to ensure sustainability.
When language meets Large Language Models
Africa’s intelligentsia has debated the place of language in education and society for decades. Kenyan author Ngũgĩ wa Thiong’o argued that the use of foreign languages in African countries could be a “cultural bomb” that “annihilates people’s belief in their names, in their languages”. Chinua Achebe held a contrary view, advocating for the use and adaptation of English by African writers.
Despite ideological differences, both wa Thiong’o and Achebe highlighted something important about Africa’s relationship with language: It is both a mirror through which people reflect on individual and communal experiences, and a projector through which they cast their beliefs and aspirations onto the world.
But what happens to these sensitivities when a large language model enters the picture?
In Guinea-Bissau, the government, with support from the World Bank Group, designed an experiment to leverage AI to reduce the cost of a large-scale learning assessment. The plan was simple: utilize Claude, Anthropic’s LLM chatbot, to co-develop assessment questions and use an AI solution to support grading and data processing nationally. The team of teachers, psychometricians and policy-makers expected that by “feeding” Claude with the national curriculum and additional context data, the chatbot would churn out highly relevant questions for the assessment. They soon discovered that the quality of results varied significantly depending on the language used.
Portuguese prompts returned lower-quality results than English prompts. Additionally, results were found lacking in terms of cultural relevance: AI-generated content would feature surnames and other words that were uncommon in the country. For instance, the content included the European Portuguese word “aldeia” for village, instead of “tabanka”, which is the common term in Bissau-Guinean Portuguese. This experience aligns with the most recent evidence, which finds persistent English–low-resource language disparities in LLMs.
The team never intended to use chatbot output verbatim, and these inaccuracies were quickly resolved. However, there had been discussions about using an AI chatbot to support teacher professional development for 6,000+ teachers nationally. A solution that suggests culturally irrelevant content to teachers might have adverse consequences on students’ learning and sense of identity. This points to the nexus between “technological” and “cultural” infrastructure, and the need for training LLMs and other AI solutions on under-represented cultures and languages.
Norms, policies and AI integration in education
In 1994, academics David Tyack and William Tobin coined the phrase “the grammar of schooling” to describe “structure and rules that organize the work of instruction”. They asserted that education technology innovations do not happen in a vacuum; norms and policies, such as those that define how students are grouped and lessons scheduled, can shape the success of education innovations. This relationship, however, is neither simple nor linear; policies can either stall or enable AI innovation in schools.
In Seychelles, the government and the World Bank Group designed an open-source AI tutor to improve secondary students’ math and geography competency. From an infrastructure perspective, conditions were optimal; Seychelles already provides free internet access to all secondary schools nationwide. However, despite ample connectivity, the implementation team struggled to find space for the pilot within the school timetable. Even when infrastructure is available, introducing regular AI tutoring sessions requires schools to decide what other activity will be displaced, who will supervise the sessions, and how participation will be sustained.
Prior to the popularization of generative AI, the Beijing Consensus called on countries and stakeholders to “adjust existing regulatory frameworks or adopt new ones to ensure responsible development and use of AI tools for education and learning”.
So how could AI policies better support the integration of the technology in education?
The World Bank Group's Digital for Tertiary Education Program (D4TEP) investigated AI and technology policies and practices at 46 South African higher education institutions. While South Africa had no national AI policy at the time, the study showed that about half of the institutions had developed their own policies and guidelines to support the ethical use of AI by students, faculty and staff.
The data also revealed that institutions that had AI policies and guidelines were substantially ahead in terms of AI deployment; among them, 78% were integrating AI into the curriculum, roughly three times the rate observed in institutions without a policy. Additionally, larger institutions were more likely to integrate AI than smaller peers. But these insights do not conclusively demonstrate that fragmented AI policies can automatically underpin the mainstreaming of the technology in education systems. Further research is needed to examine the nexus between AI policies, institutional size/resources, and the adoption and use of AI technology in education.
Transforming education systems has arguably never been more urgent. Around 1.2 billion young people in emerging and developing economies will reach working age between 2025 and 2035. Education systems should be reformed to equip youth with the right skills to find decent jobs. AI can be part of the solution. However, AI should be viewed for what it is: a tool within a broader toolkit of solutions. Reformers should refrain from placing technology systems above human systems. Policy, culture and people should be core considerations when deploying AI-powered solutions in education.





