AI for Development: Early Evidence, Risks, and Research Priorities

Artificial intelligence — machine learning, deep learning, natural language processing, computer vision — arrived in the ICT4D discourse later than mobile money or e-government, but with enormous ambition. Development organizations, technology companies, and research funders began significant investment in “AI for development” (AI4D) programs around 2017–2018, inspired by AI’s demonstrated performance in specific high-income-country applications (medical imaging, language translation, recommendation systems) and hoping to apply similar capabilities to development challenges.

The AI4D field is young and the evidence base is thin. But a growing ICT4D literature is beginning to examine what AI applications in development contexts actually do, for whom, and with what effects — including effects that were not intended.

The AI4D Landscape

AI4D programs span a wide range of applications and sectors:

Diagnostic imaging in health: Deep learning models applied to chest X-ray images can identify tuberculosis, pneumonia, and other conditions with accuracy comparable to experienced radiologists. In contexts with severe radiologist shortages, this suggests genuine opportunity to expand diagnostic capacity.

Satellite-based crop monitoring: Machine learning models trained on satellite imagery can estimate crop yields, detect crop stress, and predict harvest outcomes — at scales and speeds that field surveys cannot match. These capabilities are being used by agricultural insurance programs, food security monitoring systems, and commodity traders.

Natural language processing for local languages: NLP models — translation, speech recognition, text classification — have historically worked poorly for African and other low-resource languages that are underrepresented in training data. Investment in building datasets and models for these languages is growing, with potential applications in voice-based information services and content localization.

Humanitarian response: AI applications in humanitarian contexts include satellite imagery analysis for disaster damage assessment, population displacement tracking through mobile data, and needs assessment optimization for aid distribution.

Financial credit scoring: Alternative credit scoring models — trained on mobile money transaction history, airtime usage patterns, and other non-traditional data — have been deployed by fintech companies in Kenya, Nigeria, Ghana, and other countries to extend credit to people who lack formal credit histories.

Early Evidence: What We Know

Medical AI

The medical AI evidence base is more developed than most other AI4D application areas. Multiple studies have demonstrated that deep learning models can detect tuberculosis in chest X-rays, identify diabetic retinopathy in fundus images, and flag potential cervical cancer in colposcopy images — with performance competitive with specialist physicians.

However, the evidence on deployed-at-scale real-world performance is much thinner. Studies conducted in controlled research settings often perform better than systems deployed in actual clinical workflows with real-world image quality, equipment variation, and case mix differences. The “research to deployment gap” is significant.

A specific challenge in low-income contexts: AI medical systems are often trained on datasets from high-income countries. When applied to patients with different disease burden profiles, different comorbidity patterns, and different demographics, performance can degrade substantially.

Agricultural AI

Satellite-based crop monitoring has demonstrated reasonable accuracy for yield estimation in well-resourced contexts with good ground truth data. Performance in sub-Saharan African agricultural contexts — where smallholder plot sizes are small (often under 1 hectare), crop type diversity is high, and ground truth data for training is scarce — is less well characterized.

The most credible AI agricultural applications in the Global South to date are in food security monitoring (World Food Programme’s use of satellite data for crop stress detection) and index insurance (where satellite rainfall indices and crop growth models determine payouts). These are population-level applications where aggregate accuracy matters more than farm-level precision.

NLP for Low-Resource Languages

Progress on NLP for African languages has accelerated significantly since 2018, driven by the Masakhane community — a research collective initiated by African researchers to build NLP datasets and models for African languages. Masakhane has published datasets and models for dozens of African languages that previously had essentially no NLP resources.

Practical applications of this NLP capacity are still emerging — voice-based agricultural information in Swahili, Amharic, or Hausa; content moderation for social media in African languages; machine translation for official documents and health information. The research capability now exists; deployment at scale remains limited.

Credit Scoring

Alternative credit scoring using mobile data has been deployed at commercial scale. In Kenya, mobile credit products including M-Shwari and KCB M-Pesa use transaction history and airtime usage patterns to score creditworthiness for micro-loans. These products have reached millions of people without formal credit histories.

Research on their effects is mixed: access to credit has expanded, but over-indebtedness has also emerged as a concern — particularly for products with very short repayment periods (14-day loans) and high effective interest rates. The credit expansion benefit must be weighed against the over-indebtedness harm.

Distinctive Risks of AI in Low-Income Contexts

ICT4D researchers have identified several AI risks that are particularly significant in low-income country contexts:

Training Data Bias

AI systems learn from training data. When training data reflects historical inequality, the AI system will encode and potentially amplify that inequality. In high-income-country contexts, this has been documented in facial recognition (lower accuracy for darker skin tones trained on predominantly light-skinned datasets), hiring algorithms (discriminating against women), and criminal justice risk assessment (disparate impact on racial minorities).

In development contexts, training data biases can disadvantage the populations that development programs are meant to serve. Medical AI trained primarily on North American or European patient data may perform poorly on patients with different disease profiles or demographics. Credit scoring models trained on formal economy data may disadvantage people in informal economies in ways that are correlated with poverty.

Infrastructure Dependency

AI systems typically require reliable internet connectivity, adequate device hardware, and stable power — infrastructure that is unreliable in many low-income contexts. AI applications that work well under good infrastructure conditions may fail or degrade significantly in the actual deployment environments of low-income countries.

Accountability and Recourse

When an AI system makes a consequential decision — denying a loan, flagging a medical image as normal when it is not, misidentifying a displaced person — who is accountable? In high-income countries, accountability frameworks for algorithmic decisions are still developing. In low-income contexts with weaker institutional and legal infrastructure, accountability for AI errors is even more uncertain.

People harmed by AI decisions — denied credit, misdiagnosed, incorrectly excluded from a social program — need recourse mechanisms. These are often absent.

Labor Displacement

AI-enabled automation displaces workers. In low-income countries, the labor displaced may be workers with very limited alternative income opportunities. ICT4D researchers have begun examining the distribution of automation’s benefits and harms across income groups in developing country contexts.

Research Priorities for AI4D

The ICT4D research community has identified several priority areas for AI4D scholarship:

Context-specific performance evaluation: AI systems should be evaluated in the actual contexts where they will be deployed, not just in research settings. This requires investment in evaluation infrastructure (including ground truth datasets) in deployment contexts.

Algorithmic fairness in development contexts: Developing country adaptations of algorithmic fairness methods — which have been developed primarily in high-income country regulatory contexts — are needed. Fairness metrics appropriate for caste, ethnic, and gender disparities in specific cultural contexts may differ from those used in Western contexts.

Community engagement in AI design: Applying the participatory design tradition to AI system development for development contexts — involving intended beneficiaries in what the system optimizes for, what errors are acceptable, and what recourse mechanisms should exist.

Data sovereignty: As AI creates new value from data about people in developing countries, questions about who owns and benefits from that data are urgent. Data governance frameworks that protect community interests in their own data are an active policy and research area.

Frequently Asked Questions

Is AI going to create more inequality or less in developing countries? The honest answer is: it depends on design, regulation, and policy choices. AI’s effects on inequality are not predetermined by the technology. Well-governed AI with attention to inclusion and fairness can reduce inequality; AI deployed without these considerations is more likely to amplify existing inequalities.

What is Masakhane? Masakhane is a grassroots research initiative that emerged from African NLP researchers in 2018, aiming to build NLP data and models for African languages that have been neglected by the global NLP research community. It has grown to include hundreds of researchers across Africa and internationally, and has produced datasets, benchmarks, and models for dozens of African languages.

How does the World Bank approach AI4D? The World Bank has published guidance on responsible AI in development, invested in AI4D programs through its Digital Economy portfolio, and contributed to AI ethics frameworks. The Bank’s position is that AI creates significant opportunities for development while requiring robust governance frameworks to manage risks.

Are there privacy regulations protecting people in developing countries from AI misuse? Unevenly. Several African countries have passed data protection laws in recent years — Kenya, South Africa, Rwanda, Nigeria, Uganda — with varying scope and enforcement. Many countries still lack comprehensive data protection frameworks. The patchwork of protection leaves significant governance gaps.

What is the OECD’s AI Principles and their relevance to developing countries? The OECD adopted AI Principles in 2019 covering transparency, accountability, security, and inclusive growth. These principles are designed for OECD member countries (primarily high-income), but have been adopted by many non-member countries as a framework. The OECD has also published research on AI and the future of work in developing economies.

Development Informatics