Generative AI, Emerging Markets, and Digital Inequality

Generative AI could widen existing inequalities if access to connectivity, compute, skills, relevant data, and productive applications remains concentrated in better-prepared countries and firms. The evidence points to a dual challenge: emerging economies may face disruption in exposed occupations while lacking the infrastructure and capabilities needed to capture comparable productivity gains.^[1]^[2]

This policy synthesis examines four connected issues: multidimensional access gaps, underrepresentation of local languages and cultures, employment displacement versus opportunity, and an inclusive innovation agenda. The central conclusion is that AI policy should be treated as a development and inclusion strategy, not only as a technology-adoption programme.^[3]^[4]

1. Access Gaps Are More Than an Internet Divide

World Bank evidence identifies four foundations of meaningful generative AI participation: connectivity, compute, context, and competency. Individual adoption is growing quickly in middle-income countries, but business and government adoption remains at an early stage.^[5]

Connectivity and affordability: Internet coverage is expanding, including through technologies such as satellite connectivity, but countries continue to differ substantially in affordability, connection speed, and data usage. Reliable electricity is also necessary for affordable, AI-ready connectivity.^[6]
Compute: Access to sufficient computing capacity remains distinct from being connected. Countries and organizations may have internet access but lack the infrastructure needed to develop, adapt, or run advanced models.^[7]
Competency: Digital skills are increasingly required across occupations, while generative AI skills are spreading beyond information and communications technology roles. Low- and middle-income countries face shortages of quality, industry-aligned training and difficulties retaining talent because of brain drain.^[8]
Context: Models require large, high-quality, diverse data and must be adaptable to local economic, cultural, and institutional settings. English dominance in training data makes localization more difficult.^[9]

The policy implication is that connectivity alone is an inadequate measure of inclusion. A person may be connected but unable to afford sustained use; an organization may afford access but lack compute, skills, or locally relevant data. These layers should therefore be measured together when governments assess AI readiness.^[10]

2. Language and Cultural Representation

UNESCO describes a structural imbalance in which many AI-powered educational systems used in Africa are built, shaped, and trained around Western languages, cultures, and realities rather than local contexts.^[11] This is an access problem in a broader sense: a system that cannot reliably understand a user’s language or knowledge framework offers unequal practical access even when the user has a device and internet connection.

ChatGPT reportedly recognizes only 20% of written Hausa sentences, despite Hausa being spoken by more than 80 million people in Nigeria.^[12]
An initiative translating academic texts into Indigenous South African languages performed very poorly for Zulu, which UNESCO links largely to the limited availability of online Zulu text, given the historical importance of oral education.^[13]
GhanaGPT, which supports English and local languages including Twi, remains prone to fabricated translations and failures to interpret some concepts because Indigenous languages are insufficiently represented in training data.^[14]
Systems may also impose inappropriate cultural assumptions. For example, ChatGPT and Google Gemini answered that there are four seasons, although West Africa primarily has wet and dry seasons.^[15]
The gap includes culturally embedded knowledge, not only translation. A Kenyan teacher reported that an AI learning application could not explain a local proverb about trees.^[16]

These limitations can push learners toward English and weaken the cultural frameworks through which knowledge is interpreted.^[17] Inclusive AI therefore requires investment in local-language datasets, community validation, oral and non-text knowledge, and evaluation benchmarks that test cultural as well as linguistic accuracy. The cited evidence demonstrates the problem clearly, although it does not by itself establish how widespread each failure is across all models or countries.

3. Economic Displacement Versus Opportunity

The ILO finds that developing and emerging economies generally have lower aggregate exposure to generative-AI-driven automation than advanced economies, but they may experience disruptive effects before they can capture comparable productivity gains.^[18] Its 2025 refined index estimates that 11% of employment in low-income countries is exposed to generative AI, compared with 34% in high-income countries; globally, one in four workers are in occupations with some exposure, while 3.3% of global employment is in the highest exposure category.^[19]

Outright replacement is not the most likely outcome for most jobs. Because occupations combine tasks that can be assisted by generative AI with tasks requiring human input, job transformation is more likely than total automation.^[20] Clerical work remains the most exposed, while increasingly digitized professional and technical occupations are also showing rising exposure.^[21]

Exposure estimates require caution. Conventional measures can overstate developing-country exposure because they assume workers perform the same tasks everywhere. Skills-survey evidence indicates that workers in developing countries perform substantially fewer non-routine analytical tasks, which are the main targets of generative AI, even within occupations classified as highly exposed.^[22]

The principal inequality risk is therefore asymmetric transition rather than uniform mass unemployment. Workers in roles vulnerable to automation may have enough connectivity to experience displacement, while workers who could benefit from AI augmentation face infrastructure gaps.^[23] Country-specific exposure measures, social dialogue, and targeted employment policies are needed to manage this transition.^[24]^[25]

4. Inclusive Innovation Policy Priorities

UNDP recommends a people-centered policy package that expands national readiness while treating AI as a shared public good rather than a concentrated advantage.^[26] The priorities below connect the access, representation, and employment findings.

Make access affordable and usable: Expand reliable connectivity, electricity, devices, and data access, with particular attention to lower-readiness countries and underserved rural communities. UNDP reports that AI use remains close to 5% in many low-income countries, compared with two in three people in some high-income economies.^[27]^[28]
Build broad-based skills: Fund AI fluency and digital skills through schools, vocational systems, worker retraining, and public institutions. Programmes should prioritize women, rural communities, minorities, and other groups less likely to have access to digital tools.^[29]^[30]
Support local-language and community-led innovation: Invest in representative datasets, local-language models, oral knowledge documentation, and testing by educators and communities. Data collection and design should make visible whose needs and experiences are missing.^[31]
Develop sustainable compute and public digital infrastructure: Support affordable, energy-conscious compute and shared infrastructure so local firms, universities, and governments can adapt systems to national needs rather than remaining only consumers of imported tools.^[32]^[33]
Protect workers through managed transition: Use country-specific occupational analysis, social dialogue, worker consultation, retraining, and employment services to address task transformation and unequal exposure.^[34]^[35]
Govern for accountability and public value: Establish safeguards against opaque systems, misuse, data exclusion, misinformation, and security risks. Direct innovation toward health, education, agriculture, transport, financial inclusion, public services, and climate and disaster resilience.^[36]^[37]^[38]
Cooperate regionally and globally: Coordinate standards, safety practices, and open models so AI capability does not become concentrated in a small number of countries or firms.^[39]

Implementation should be nationally tailored. The World Bank emphasizes that the appropriate balance among connectivity, compute, context, and competency depends on each country’s circumstances, while UNDP calls for strategies aligned with national capacity.^[40]^[41] Governments should track not only adoption rates, but also affordability, local-language performance, distribution of skills, worker outcomes, and whether public-interest applications reach underserved groups.

Executive Conclusion

Generative AI is likely to produce both opportunity and inequality in emerging markets. The strongest evidence does not support a simple story of either universal job loss or automatic development gains. Instead, it shows layered access barriers, serious language and cultural gaps, and a risk that disruption will reach some workers before augmentation reaches those with the greatest infrastructure constraints.^[42]^[43]^[44]

The immediate policy priorities are to close affordable access gaps, invest in skills and sustainable compute, build locally representative language resources, protect workers through social dialogue and transition support, and govern AI toward accountable public-interest uses. This package offers the clearest route to reducing the risk of a widening divergence between countries that can capture the AI dividend and those that face exclusion from it.^[45]^[46]