The Literacy Trap

Between 2022 and 2024, job vacancies requiring generative AI skills surged ninefold globally. In lower-middle-income countries, AI-related job postings grew at 11 percent, faster than in high-income economies, where the growth rate was just 2 percent. The World Bank, which published these figures in its 2025 Digital Progress and Trends Report, presented them as evidence of opportunity. But read alongside another number in the same report — that lower-middle-income countries account for just 5 percent of global data centre capacity, and low-income countries for less than 0.1 percent, they also describe something more complicated than opportunity. What they actually describe is a workforce being prepared to participate in an infrastructure it does not own, govern, or in most cases meaningfully influence.

This is the context in which the development sector's AI literacy programmes are being built. And it is a context that most of those programmes are not yet designed to address.

By 2025, “AI-readiness” had become the new shorthand for progress, much like “digital transformation” before it — woven into grant requirements and organisational capacity assessments. NGOs and CSOs across Africa, South Asia, and Southeast Asia found themselves required to demonstrate AI fluency to remain competitive for funding, regardless of whether AI was relevant to their programmatic priorities. The technology integration came first. Strategy struggled to catch up, and training often followed as a way to make sense of it after the fact.

This sequencing has a structural consequence: when adoption is the implicit goal, programmes optimise for tool competence; “readiness” simply becomes a compliance requirement and not really a design imperative.

This is what cognitive scientist Abeba Birhane, writing in SCRIPTed, identified as algorithmic colonisation: the importation of AI tools built in the West, for Western contexts, with Western incentive structures, which simultaneously "impoverishes development of local products while leaving the continent dependent on Western software and infrastructure." This creates a circularity worth naming: the funding conditions that accelerated AI adoption in the sector are largely the same ones now funding programmes to manage its consequences.

The AI literacy curricula now circulating across low- and middle-income countries (LMICs) are largely inherited from the digital literacy frameworks of the 2010s. Those frameworks were designed for a world of tools: browsers, search engines, social media platforms. Their underlying logic was about individual access and agency: “Can a person find, evaluate, and act on information in a digital environment?”

AI systems operate on a different logic. They are not primarily tools that individuals use. They are systems that make, influence, or shape decisions, and most often about the very individuals who encounter them. Consider a credit scoring model, an agricultural advisory system, or a content moderation tool. These are all decision-making infrastructures that determine outcomes. Often at population scale. Often without clear explanation. And often without mechanisms that allow those affected to question, contest, or even fully understand what has been decided on their behalf.

The cognitive and civic demands this creates are categorically different from what digital literacy was originally built to address. 

Digital literacy asks: can you use this? AI literacy, done properly, asks something harder: can you interrogate this? Can you ask what it is optimising for, whose interests shaped its design, what it gets wrong and for whom, and what recourse exists when it fails? These are as much technical questions as they are ethical ones. They demand different forms of knowledge, different pedagogical approaches, and definitely different ways of defining and measuring outcomes.

The trap manifests consistently across contexts.

The competent user who cannot contest. Participants leave programmes with genuine, if narrow, capability: they can prompt effectively, flag hallucinations, apply an ethics checklist, and produce AI-assisted outputs. What they leave without is contestability. No framework for challenging an AI-driven decision that affects them, no vocabulary for advocacy, no understanding of what redress looks like in regulatory environments where the right to explanation does not exist in law. In much of sub-Saharan Africa, AI is deployed by governments with minimal transparency obligations. For instance, the OECD/EC’s draft AI Literacy Framework equips learners to engage with AI in contexts where data protection law exists, regulatory bodies enforce it, and recourse is accessible. In most contexts where development sector programmes operate, those conditions are absent or nascent, which means the competencies the framework builds have nowhere to land.

Frameworks that travel poorly. The responsible AI principles embedded in most multilateral toolkits were designed for contexts with functioning data protection laws, independent judiciaries, and competitive AI markets. As Birhane's work on algorithmic colonisation makes clear, the importation of frameworks built for one regulatory ecology into another does not produce safety, but only the performance of safety. 

The alternative is already being constructed, unevenly, in the regions the sector claims to serve.

Current programmes position the learner as a user. As someone to be equipped with skills to operate systems designed elsewhere, under terms set by others. A more adequate approach would position the learner (and crucially, the community), as a subject with civic standing. Someone with both the right and the capacity to participate in decisions about whether and how AI enters their ecosystems.

This reframing has concrete pedagogical implications. It means building interrogative capacity alongside tool competence: how to ask who built this, on what data, with what incentives, and who bears the cost of its errors. It means designing for collective sense-making, because AI systems affect communities at scale and not individuals in isolation. And it means orienting programmes explicitly toward governance participation, building the knowledge and confidence for communities to engage with the institutions, however weak or distant, that govern AI deployment in their contexts.

None of this is starting from zero. The AU Strategy, Masakhane, and the CARE Principles each point toward what a governance-first approach looks like when it is taken seriously, as the foundation and not an add-on to adoption. The opportunity for the sector is to recognise the alternative approaches to AI literacy and integration already being built, and ask honestly whether its funding frameworks are designed to support them.