Chapter 34: AI Ethics in Emerging Markets

The Leapfrog and the Extraction

When Safaricom deployed M-Pesa in Kenya in 2007, it proved something that development economists had long theorized but rarely seen demonstrated so dramatically: that developing economies do not have to follow the same technological path as wealthy ones. Mobile money allowed Kenyans to send and receive payments through basic feature phones without bank accounts, physical infrastructure, or the decades of regulatory and financial institution development that mobile payments required in Europe and North America. By 2023, M-Pesa had 51 million customers. The lesson was clear: technology can leapfrog infrastructure gaps, and populations that wealthy-country institutions had written off as "unserveable" could adopt transformative technology rapidly when it was designed to serve their actual needs.

The AI equivalent of M-Pesa's promise is visible across the Global South. Agricultural AI that identifies crop diseases from smartphone photographs is helping smallholder farmers in Uganda diagnose problems that would have required expensive agronomists. AI-powered diagnostic tools are extending the reach of limited healthcare workforces in settings where doctors are vastly outnumbered by patients. Natural language processing for Swahili, Amharic, and dozens of other languages that large language models historically ignored is opening possibilities for digital inclusion that English-only AI foreclosed. The promise is real.

But the M-Pesa story also carries a darker lesson that the leapfrog narrative tends to suppress. M-Pesa's success benefited Safaricom, which is partly owned by Vodafone and a Kenyan government investment fund. The platform generated enormous transaction data on Kenyan financial behavior — data that was owned by the company, not the communities whose economic life it represented. The regulatory framework governing data use was weak and developed retroactively. International financial institutions and development finance organizations that promoted mobile money as a development tool had limited engagement with questions about data governance, competitive dynamics, or the concentration of financial infrastructure in a small number of private companies. The story of technology deployment in emerging markets is never purely a story of inclusion and empowerment — it is always also a story about who benefits from data, who builds the infrastructure, who sets the terms, and who is accountable when things go wrong.

These tensions — between promise and extraction, between leapfrogging and replication of colonial patterns, between locally appropriate innovation and externally imposed technology — define the AI ethics landscape in emerging markets. This chapter examines that landscape with honesty about both dimensions.

Learning Objectives

By the end of this chapter, you will be able to:

1. Critique the framing of "emerging markets" as a category, identifying the diversity it contains and the analytical problems that arise from treating the Global South as a monolith.

2. Identify the most significant beneficial AI applications in emerging market contexts — agricultural AI, healthcare AI, financial inclusion AI, educational AI, and language AI — and assess the conditions under which they genuinely serve local populations.

3. Analyze the concept of "data colonialism" and explain how the extraction of data from Global South populations for AI training without adequate benefit-sharing replicates historical colonial economic patterns.

4. Identify the key bias and representation gaps that affect AI systems deployed in emerging markets, including language model underrepresentation, agricultural AI trained on non-local crop varieties, and medical AI not calibrated for local patient populations.

5. Describe the pattern of surveillance technology export to authoritarian and semi-authoritarian governments in the Global South, and assess the implications for democratic governance and human rights.

6. Evaluate the relationship between digital infrastructure gaps and AI equity, explaining why AI systems that require reliable connectivity, smartphones, and electricity replicate rather than address existing inequalities.

7. Identify the most important local AI governance developments in Africa, India, Brazil, and other major Global South regions, and assess the challenges they face.

8. Apply a framework for "responsible international AI deployment" to assess whether a specific AI deployment in an emerging market context meets genuine ethical standards.

Section 1: The Framing Problem

The term "emerging markets" is itself analytically problematic, and beginning with that problem is important. Coined by IFC economist Antoine van Agtmael in 1981 as a marketing term to make investors less squeamish about "Third World" investing, the label encompasses an extraordinary diversity of economic, political, and social contexts. It encompasses India with 1.4 billion people and a sophisticated technology sector, and Bhutan with 800,000 people and a recently developed internet infrastructure. It encompasses Brazil — a G20 economy with a GDP larger than Italy's — and Chad, which remains one of the world's poorest countries by almost any measure. It encompasses the Gulf states, which have among the world's highest per capita incomes, and sub-Saharan African economies where the majority of the population lives on under $5 per day.

This diversity has profound implications for AI ethics analysis. The AI challenges facing India — where the technology sector is among the most sophisticated in the world, AI talent is abundant, and the government is actively developing AI governance frameworks — are profoundly different from the challenges facing the Democratic Republic of Congo, where internet penetration is low, regulatory capacity is limited, and AI systems are primarily experienced as imported products from outside rather than domestically developed tools. Analysis that treats these contexts as equivalent — that speaks of "the emerging market AI challenge" as if it were a single thing — will produce both intellectually impoverished analysis and practically useless recommendations.

Within sub-Saharan Africa alone, the diversity is extraordinary. South Africa has a relatively sophisticated regulatory environment, high internet penetration (by African standards), a significant domestic technology sector, and a complex post-apartheid history that makes issues of algorithmic bias and surveillance particularly politically charged. Nigeria, the continent's largest economy and most populous country, has a fast-growing startup ecosystem, a massive and underserved population, and a regulatory environment that is evolving rapidly. Somalia has almost no functional central government, minimal digital infrastructure, and experiences AI primarily through the conflict surveillance tools that international military and humanitarian organizations deploy.

A second problem with the "emerging markets" frame is its teleological implication — that these markets are "emerging" toward the wealthy country model, that the trajectory is from underdevelopment toward the US or European economic structure. This framing obscures the ways in which Global South economies have distinct economic structures, distinct cultural contexts, and distinct governance traditions that AI systems developed elsewhere may not serve well and may actively undermine. The assumption that AI governance frameworks developed in Washington, Brussels, and Beijing are appropriate for Lagos, Dhaka, and Lima is precisely the assumption that chapter aims to interrogate.

We use the "emerging markets" term throughout this chapter while being explicit about its limitations, because it is the term most commonly used in the business and policy literature this chapter addresses. Readers should understand it as a convenient shorthand for a heterogeneous collection of economies, not as a description of a coherent analytical category.

Section 2: AI's Promise in Emerging Markets

The potential benefits of AI in emerging market contexts are genuine and, in some cases, transformative. Understanding these benefits concretely — not as abstract claims about AI's potential but as documented applications with measured outcomes — is important for balanced analysis. The point of examining AI's extractive risks is not to conclude that AI has nothing to offer the Global South, but to understand what conditions are necessary for AI's promise to be realized.

Agricultural AI has perhaps the most immediate and broadly applicable development potential. Agriculture remains the primary livelihood for the majority of the world's poorest people, most of whom are smallholder farmers in sub-Saharan Africa, South Asia, and Southeast Asia. These farmers face challenges that AI can directly address: crop disease identification, where AI-powered smartphone applications can diagnose plant diseases from photographs and recommend treatments without requiring access to agronomists who are scarce and expensive; soil health analysis, where machine learning models can predict crop yield responses to different inputs; weather and climate prediction at localized scales, where AI models trained on satellite and ground station data can provide actionable agricultural advice; and market price information, where AI systems can aggregate price data across markets and provide smallholders with information that previously only large traders possessed.

The evidence that agricultural AI can produce meaningful benefits for smallholder farmers is growing. The Plantix application, developed by PEAT, has documented significant adoption among South Asian and East African farmers for crop disease diagnosis. The Hello Tractor platform uses AI to match smallholder farmers with tractor services in Nigeria and Kenya, improving mechanization access. These are not hypothetical benefits — they are documented improvements in productivity, income, and livelihoods. The conditions for realizing them, however, are demanding: the AI must be trained on locally relevant data (African crop varieties and disease pressures differ significantly from European or North American ones), accessible through connectivity and devices that smallholders actually have, and integrated into agricultural extension systems that can support adoption.

Healthcare AI offers similarly compelling potential in resource-constrained settings. In sub-Saharan Africa, there are approximately 0.2 physicians per 1,000 people — compared to 3.6 per 1,000 in the United States and 4.3 per 1,000 in Western Europe. AI diagnostic tools cannot replace physicians, but they can extend the reach of the physicians who exist, enable community health workers to identify conditions requiring referral, and support clinical decision-making in settings where specialist expertise is unavailable. AI-assisted tuberculosis screening from chest X-rays has been validated in several African contexts. AI retinal screening for diabetic retinopathy detection is active in India. AI-assisted cervical cancer screening tools, designed for low-resource settings without specialist pathology capacity, are in use across sub-Saharan Africa.

The conditions for healthcare AI to serve rather than harm in these contexts are demanding, however. Models trained on patient populations in the United States or Europe may not perform adequately on African or South Asian patient populations, where disease prevalence, comorbidity patterns, and population genetics differ. Deployment without adequate training, oversight, and integration into health systems can produce false reassurance and misdiagnosis. And the infrastructure requirements — consistent electricity, internet connectivity, compatible device ecosystems — exclude the most remote and underserved communities.

Financial inclusion AI builds on the M-Pesa model, extending credit, insurance, and savings products to people who lack the conventional credit histories, collateral, and formal employment records that traditional financial products require. Companies like Tala, Branch, and Jumo use machine learning models trained on mobile phone behavioral data — airtime top-up patterns, mobile money transaction history, app usage — to extend credit to borrowers who are "credit invisible" to conventional lenders. These approaches have genuine inclusion potential: they extend financial services to populations that formal financial institutions have excluded, often enabling small business investments and consumption smoothing that have measurable impacts on economic wellbeing.

The risks are correspondingly significant. Machine learning credit models trained on mobile phone behavior may encode discrimination against populations whose phone usage patterns differ systematically — women who use phones less than men, elderly people, people in rural areas with limited connectivity. The transparency and explainability of these models is often limited, making it difficult for borrowers to understand why they were denied credit or what they can do to improve their creditworthiness. And the concentration of financial inclusion in a small number of large platforms creates risks of data extraction, monopolistic pricing, and vulnerability to platform failures that serve neither borrowers nor the financial system.

Language AI represents a particularly important application domain because the English-language dominance of AI training has created a profound representation gap. Of the world's approximately 7,000 languages, only a small fraction are adequately represented in large language models. African languages — of which there are more than 2,000 — are dramatically underrepresented relative to their speaker populations. The consequences are not merely that AI assistants don't speak Swahili: they extend to search engines that return worse results in Hausa, medical AI that cannot communicate with patients in their preferred language, educational AI that forces students to learn in foreign languages, and government services that exclude citizens who don't communicate in official colonial languages. Community-driven AI research initiatives like Masakhane NLP are building the datasets and models needed to address this gap, with measurable progress in African language NLP.

Section 3: AI's Extractive Risks

The promise of AI in emerging markets is real but conditional. The conditions — that AI systems be designed for local contexts, trained on locally relevant data, governed by local institutions, and accountable to local communities — are not consistently met. When they are not, AI in emerging markets can replicate and intensify historical patterns of economic extraction.

The concept of "data colonialism," articulated by scholars including Nick Couldry and Ulises Mejias in their 2019 book of the same name, provides a useful framework for understanding how this extraction operates. Their argument, developed further by a growing body of critical data studies scholarship, is that the collection of data from Global South populations is analogous in important ways to historical colonialism's extraction of natural resources. Data generated by people's daily lives — their economic transactions, their social interactions, their health experiences, their agricultural practices — is collected by platforms and devices often owned by companies in wealthy countries, used to train AI systems developed by those companies, and monetized primarily to benefit shareholders who are also predominantly in wealthy countries. The populations whose data was extracted may receive some services in return, but they have minimal ownership, control, or governance power over the systems built from their data.

This extractive pattern is documented across multiple domains. The global AI training data industry relies significantly on low-cost annotation labor in countries like Kenya, Uganda, the Philippines, and India — workers who identify and label the images, text, and audio that make AI systems function. The workers who do this labor typically receive wages far below the value of the data products they create, work in conditions that can involve significant psychological harm when labeling violent or disturbing content, and have minimal labor protections or recourse when conditions are inadequate. The TIME magazine investigation of OpenAI's content moderation outsourcing to Kenyan workers through the Sama Group, detailed in Case Study 1, illustrates this dynamic with disturbing specificity.

The training data extraction problem extends beyond annotation labor. Agricultural data generated by smallholder farmers using AI-powered diagnostic apps is collected by platforms whose data governance terms are often opaque and rarely negotiated in local languages. Health data from AI diagnostic tools in African healthcare settings is often used to improve models that are then sold at prices that exclude the health systems whose data contributed to their development. Financial transaction data from mobile money platforms generates analytical insights that their corporate owners use to develop more lucrative products, with limited benefit-sharing to the communities whose economic life created the data.

The extractive pattern is not inevitable — it is a design choice. Benefit-sharing agreements can be structured to return value to data-contributing communities. Open-source AI development can enable local institutions to use and adapt AI systems without dependence on foreign corporate platforms. Data sovereignty frameworks can give national institutions meaningful control over data generated within their borders. But these choices require that extraction be recognized as the default that it currently is, rather than normalized as the natural order of technology development.

Section 4: Bias and Representation Gaps

AI systems developed primarily by researchers and engineers in wealthy, English-speaking countries, trained primarily on data generated in those countries, have systematic failures when deployed in contexts their development did not adequately consider. These failures are not random — they consistently disadvantage populations who were already marginalized.

Language model bias is perhaps the most visible representation gap. Large language models perform significantly worse in African languages, many South Asian and Southeast Asian languages, and indigenous languages globally than they do in English, French, and other dominant languages. This performance gap is not merely a nuisance — it determines whether people can benefit from AI tools at all. A medical AI assistant that cannot communicate accurately in Amharic or Yoruba cannot serve healthcare workers in Ethiopia or Nigeria. An educational AI that functions well only in English reproduces the linguistic exclusions of colonial education systems.

The causes of language model underrepresentation are structural. Large language models are trained on internet text data, and internet text data reflects the linguistic demographics of internet users — which are heavily skewed toward wealthy, educated, predominantly male populations in a small number of countries. African languages, despite having hundreds of millions of speakers, are represented by a tiny fraction of internet text because their speakers have lower internet access rates and because colonial language policies have made European languages dominant in formal written communication even where African languages dominate in daily life. Improving language model performance for underrepresented languages requires deliberate investment in creating and collecting training data in those languages — investment that is profitable primarily for companies that see commercial value in those markets.

Agricultural AI bias has direct consequences for food security. AI crop disease detection models trained primarily on images of North American and European crop varieties and disease pressures may perform poorly when applied to the varieties of those crops grown in Africa or Asia, which are different in appearance and face different disease pressures. A model trained to detect late blight on Russet Burbank potatoes in Washington State may fail to recognize late blight on the local potato varieties grown in highland Kenya, even though the underlying pathogen is the same. Similarly, AI-powered agricultural recommendation systems trained on US or European farm data may give inappropriate recommendations for smallholder farming systems with different soil types, rainfall patterns, crop rotation practices, and input availability. The consequences of agricultural AI failures are not merely inconvenient — they can directly affect food security for subsistence farmers.

Medical AI bias by population creates risks to health and life. AI diagnostic models trained predominantly on clinical data from North American and European health systems systematically underperform for patients whose disease prevalence, symptom presentation, and physiological characteristics differ from those training populations. Dermatology AI, trained overwhelmingly on images of light-skinned patients, has documented performance gaps for patients with darker skin tones. Sepsis prediction models calibrated on US hospital populations may have different performance characteristics when deployed in African hospital settings where the causes and clinical presentation of sepsis differ. Chest X-ray interpretation AI trained on North American radiology data may miss endemic diseases — tuberculosis, schistosomiasis, parasitic infections — that are common in Sub-Saharan Africa but rare in North American training populations. These are not theoretical concerns: they are documented performance gaps with direct implications for patient safety.

Section 5: Surveillance Export

One of the most troubling dimensions of AI in emerging markets is the export of surveillance technology to governments in the Global South that use it for political repression. The pattern is documented, widespread, and in fundamental tension with the democratic governance and human rights frameworks that provide the normative foundation for AI ethics.

China's export of AI surveillance infrastructure through companies like Huawei, ZTE, Hikvision, and Dahua has been the subject of extensive reporting and research. The "Safe City" programs that Huawei has marketed and deployed across Africa — including in Zimbabwe, Uganda, Ethiopia, Kenya, Angola, and Ivory Coast — package AI-enabled surveillance cameras, facial recognition systems, command and control software, and data management infrastructure into integrated urban security systems. The marketing emphasizes crime reduction and traffic management; the actual applications in several documented cases have included surveillance of political opposition, identification of protesters, and monitoring of journalists and civil society activists.

The Zimbabwe case is among the most extensively documented. Huawei supplied the Zimbabwean government with surveillance camera networks and reportedly assisted in developing facial recognition capacity under a 2018 memorandum of understanding with the Zimbabwean government's postal and telecommunications regulatory authority. Robert Mugabe's government — and subsequently Emmerson Mnangagwa's — had well-documented histories of using information about political opponents to facilitate harassment, imprisonment, and violence. The introduction of AI-enabled mass surveillance technology into this political context had predictable implications for political opposition, journalists, and civil society organizations that monitoring systems had previously only partially identified.

The Uganda case illustrates the technology's direct political application. The Ugandan government, which has ruled under President Yoweri Museveni since 1986, used Huawei-supplied surveillance technology as part of a broader effort to monitor political opposition leading up to the 2021 elections. Reports documented the use of surveillance infrastructure to track opposition politician Bobi Wine (Robert Kyagulanyi) and his supporters. The technology did not cause the political repression — Uganda's human rights violations long predate Chinese technology exports — but it provided tools that made surveillance more systematic, scalable, and effective.

The pattern is not limited to Chinese technology exports. US-origin surveillance technology has also been sold to authoritarian governments in the Middle East and Southeast Asia. Israeli surveillance company NSO Group's Pegasus spyware has been documented in use by governments including Saudi Arabia, the UAE, Morocco, Rwanda, and India against journalists, human rights lawyers, and political opponents. The export of powerful surveillance technology without adequate governance requirements — requirements that it not be used against political dissidents, journalists, or civil society — has been documented as a significant contribution to democratic backsliding globally.

The governance failures are multiple: inadequate export controls by technology-exporting governments; insufficient due diligence by technology companies about how their products are used; absence of end-use monitoring; and limited accountability when misuse occurs. The AI ethics implication is direct: organizations involved in developing or deploying surveillance AI must grapple seriously with how their technology may be used by governments with poor human rights records. "We didn't know how it would be used" is not an adequate response when the risks are foreseeable.

Section 6: Digital Infrastructure Gaps

A fundamental constraint on AI's potential in the Global South is the infrastructure it requires. Most current AI applications — particularly those that rely on large language models, cloud inference, or real-time video analysis — require reliable internet connectivity, capable smartphones, and stable electricity supply. None of these can be assumed across much of the Global South, and the populations most likely to be left behind by AI's advance are precisely those with the least access to this infrastructure.

Internet penetration rates vary enormously. As of 2024, internet penetration in sub-Saharan Africa averages around 36%, with enormous variation across countries (South Africa at approximately 68%, Ethiopia at approximately 25%, Central African Republic under 10%). In South Asia, India's rapid mobile internet expansion has reached approximately 50% of the population, but access is unequally distributed between urban and rural areas, between men and women, and between economic classes. The populations with the lowest internet access are typically the same populations that face the greatest development challenges and would in principle benefit most from AI applications designed to serve their needs.

Smartphone access creates an additional filter. While feature phones are nearly universal in many African and South Asian markets, AI applications that depend on significant processing power, large screen displays, or camera functionality require smartphones that remain inaccessible to significant portions of low-income populations. Agricultural AI apps that identify crop diseases from photographs require a smartphone camera of adequate quality — not universally available among the smallholder farmers who could most benefit from crop disease diagnosis.

Electricity supply is a third critical infrastructure gap. Many African countries have electricity supply that is unreliable (power cuts averaging several hours per day are common in countries including Nigeria, Ghana, Tanzania, and Ethiopia), unavailable (only 43% of sub-Saharan Africa's population had access to electricity as of 2022, with most of the unconnected in rural areas), or unaffordable. AI systems that require continuous internet connectivity — for cloud inference, model updates, or data synchronization — simply cannot function reliably in these contexts.

The infrastructure gap creates an equity paradox: the populations with the greatest development needs are systematically less likely to benefit from AI applications designed without adequate attention to infrastructure constraints. This is not inevitable — AI applications designed specifically for low-resource environments (small models that can run on feature phones, offline functionality, compatibility with 2G rather than 4G connectivity) can serve populations that would otherwise be excluded. But designing for these constraints requires deliberate choice, and the commercial incentives typically point toward building for more lucrative, better-connected markets.

Section 7: Local AI Governance Development

The development of domestic AI governance capacity in Global South countries is both necessary and ongoing. Understanding what these frameworks look like, what challenges they face, and where they are succeeding provides important context for organizations operating in these markets.

African governance development has proceeded at multiple levels. The African Union's Digital Transformation Strategy for Africa, adopted in 2020, and the subsequent work on the AI Continental Strategy described in Chapter 32 represent the continental level. At the national level, several African countries have developed or are developing national AI strategies and governance frameworks. Rwanda's National AI Policy, adopted in 2019, is among the continent's more sophisticated — reflecting the Rwandan government's broader technology-centered development strategy. Kenya's Digital Economy Blueprint includes AI governance elements. Egypt has developed an AI strategy with specific governance components. Nigeria's National AI Policy (2021) sets out principles and implementation priorities. South Africa's AI Institute is advancing both research and policy development.

The common challenge across these frameworks is the gap between policy aspiration and implementation capacity. Developing sophisticated AI regulation requires technical expertise in machine learning, data science, and AI system design that is in genuinely short supply in most African countries, concentrated in a small number of universities and companies, and subject to significant brain drain toward international employers who offer substantially higher compensation. Building regulatory capacity requires sustained investment in training regulators, developing technical guidance resources, and building the institutional infrastructure needed for enforcement.

India's approach has been notably cautious about imposing regulation, reflecting the government's priority on making India a global AI development hub and leveraging India's significant comparative advantage in AI talent and software development. India's Digital Personal Data Protection Act (2023) applies to AI processing of personal data. India's AI governance more broadly has relied on voluntary principles issued by agencies like NASSCOM and NITI Aayog rather than comprehensive legislation — a contrast with the EU's more prescriptive approach that reflects India's economic development priorities.

Brazil's development combines a GDPR-inspired data protection law (LGPD, 2018) with ongoing legislative development of a more comprehensive AI bill. Brazil's Constitution's provisions on human dignity, equality, and access to technology have provided a constitutional framework within which AI governance is developing. Brazil's AI governance discussions are relatively sophisticated, reflecting both the country's significant domestic AI sector and its civil society's engagement with digital rights issues.

Section 8: The Representation Problem in Global AI

The governance gap identified in Chapter 32 — the systematic underrepresentation of Global South nations, perspectives, and communities in global AI governance — has a mirror image in the AI systems themselves. AI systems are not neutral tools; they encode the values, assumptions, and worldviews of the people and institutions that build them. When AI is built predominantly by researchers and engineers in a small number of wealthy, English-speaking countries, it inevitably encodes those contexts' assumptions in ways that may not serve — and may actively harm — people in other contexts.

The demographic homogeneity of the AI field is a documented problem. Research from leading AI conferences and labs consistently shows that the AI research community is overwhelmingly male, overwhelmingly from a small number of countries (the US, China, and a handful of European nations dominate AI publication), and lacking in racial and ethnic diversity even within wealthy countries. A 2023 analysis of faculty at top AI programs found that less than 5% were Black and less than 5% were Hispanic, even at US universities. The Global South is even more dramatically underrepresented: the continent of Africa, with 1.4 billion people, produces a tiny fraction of global AI research output.

This concentration of AI development has consequences that go beyond symbolic representation. The research agenda of the AI field is shaped by the priorities of its dominant institutions. Applications relevant to wealthy-country consumers and governments — natural language processing for English, image recognition for autonomous vehicles on Western road systems, AI for financial trading — receive enormous research attention. Applications relevant to Global South populations — AI for local language preservation, agricultural AI for smallholder farming systems, AI for disease contexts endemic to tropical regions — receive minimal attention relative to their potential impact.

The values embedded in AI systems also reflect their development context. What counts as a "fair" AI outcome, what information sources are treated as authoritative, what social relationships and economic arrangements are treated as normal, what political systems are treated as the baseline — all of these reflect the cultural and political assumptions of the development environment. A credit scoring model trained on US credit market data encodes assumptions about employment patterns, banking behavior, and creditworthiness signals that may not translate accurately to other economic contexts. A content moderation system trained on US political and cultural norms may suppress speech that is normal and valuable in other political contexts while being insensitive to harms specific to other cultures.

Section 9: Genuine Partnership Models

The extractive model of AI deployment in the Global South is not the only model available. Several initiatives demonstrate what genuinely reciprocal, locally-driven AI development can look like.

Masakhane NLP is perhaps the most compelling example of community-driven AI research for African contexts. Founded in 2019 by South African NLP researcher Jade Abbott and a growing community of African researchers, Masakhane has built machine translation and language processing capabilities for more than 100 African languages through a distributed research model that centers African researchers, institutions, and communities rather than treating Africa primarily as a source of linguistic data for research conducted elsewhere. Masakhane's "participatory research" approach explicitly addresses data sovereignty — African language data collected for Masakhane projects remains in African institutions — and its open publication model ensures that research outputs are accessible to African researchers and developers without requiring access to expensive proprietary systems. The initiative demonstrates that high-quality AI research on African contexts can be driven by African researchers with appropriate institutional and funding support.

FAIR for Africa represents a major tech company's attempt to build research capacity in Africa rather than simply extract data from it. Meta's FAIR (Fundamental AI Research) program established a Nairobi-based research lab that employs African AI researchers to work on problems relevant to African contexts, including African language NLP, agricultural AI, and AI for healthcare in low-resource settings. The initiative is not without critics — concerns about whether corporate research labs, however locally staffed, can genuinely serve local interests rather than corporate ones are legitimate — but it represents a different model than simply deploying products designed elsewhere.

Data cooperatives and sovereignty frameworks offer another model for more equitable AI development. The concept of a data cooperative — an organization that aggregates data from its members, negotiates collective data sharing agreements, and ensures that members benefit from the value their data creates — has been explored in several African contexts. Agricultural data cooperatives, in which smallholder farmers collectively own and govern the data generated by their farming practices, could give farmers collective bargaining power over how their data is used and ensure they share in the value it creates. These models remain early-stage but represent important experiments in alternatives to the extraction model.

Section 10: What Organizations Should Do

For business professionals operating in emerging market contexts, the AI ethics considerations discussed in this chapter translate into practical responsibilities that differ from and extend beyond the compliance obligations discussed in Chapter 33.

Responsible international AI deployment begins with genuine understanding of the context into which AI is being deployed. Organizations deploying AI systems in African, South Asian, Southeast Asian, or Latin American markets should invest in understanding the local context — the infrastructure constraints, the regulatory environment, the cultural norms, the existing governance institutions, and the communities affected — before deployment. This understanding cannot be achieved from headquarters; it requires engagement with local partners, communities, and experts.

Community engagement is a non-negotiable element of responsible deployment in contexts where communities have historical experience of external entities extracting value without adequate return. Community engagement means more than market research about product-market fit. It means genuine consultation with affected communities about whether and how they want AI systems to operate in their lives, what governance frameworks they consider adequate, and what benefit-sharing arrangements they consider fair. This consultation must happen before deployment, must genuinely influence deployment decisions (not merely be documented as having occurred), and must include communities that are affected even if they are not customers.

Data sovereignty agreements provide a concrete mechanism for operationalizing benefit-sharing commitments. Organizations that collect data from Global South communities for AI training should negotiate explicit agreements about data ownership, governance, use limitations, and benefit distribution — agreements that are legally enforceable by community institutions, negotiated in local languages, and structured to protect communities' ongoing interests rather than merely satisfying disclosure requirements at the point of data collection.

Local partnership requirements — genuine co-development and co-governance rather than vendor relationships with local resellers — create accountability structures that serve communities better than purely extractive models. When AI systems are co-developed with local institutions that have relationships with affected communities, those institutions can advocate for community interests within the development process, adapt systems to local contexts, and provide ongoing accountability mechanisms that remote corporate headquarters cannot provide.

Avoiding the extraction model means actively designing against patterns that replicate historical colonial dynamics: ensuring that data collected in Global South contexts primarily benefits the communities from which it was collected; building local AI capacity rather than creating long-term dependence on foreign platforms; supporting local governance frameworks rather than structuring operations to avoid local regulatory oversight; and pricing AI products in ways that make them genuinely accessible to local institutions and communities rather than extracting maximum value for global shareholders.

These responsibilities are not merely ethical aspirations — they are increasingly becoming commercial requirements. As AI regulation develops across emerging markets, organizations that have built extractive business models will face growing regulatory risk. As communities develop greater awareness of data rights and greater capacity to advocate for them, the reputational costs of extraction will increase. And the organizations that build genuine trust with Global South communities and governments will have durable competitive advantages in the world's fastest-growing markets.

Summary

AI in emerging markets presents both genuine promise and serious ethical risk. The promise — agricultural AI for smallholder farmers, healthcare AI for resource-constrained health systems, financial inclusion AI for the unbanked, language AI for underrepresented languages — is real and documented. But it is conditional on AI systems being designed for local contexts, trained on locally relevant data, governed by local institutions, and accountable to local communities.

The risks are equally real: data colonialism that extracts value from Global South data without adequate benefit-sharing; bias and representation gaps that make AI systems less effective and potentially harmful for non-Western populations; surveillance technology export that enables authoritarian repression; and digital infrastructure gaps that exclude the most vulnerable populations from AI's benefits. These are not hypothetical futures — they are documented present realities.

Addressing them requires moving beyond the extractive model of AI deployment — in which technology developed in wealthy countries is deployed in the Global South primarily to serve the financial interests of its developers — toward genuine partnership models that center local communities, build local capacity, and ensure that AI's benefits are shared by the communities whose lives and data make it possible.

This chapter continues with Case Study 1: Content Moderation and the African Labor Force, Case Study 2: Huawei's Safe City Programs in Africa, Key Takeaways, Exercises, Quiz, and Further Reading.