General
Data Labeling in the Global South
Published March 11, 2026
The AI industry runs on a workforce it would prefer you didn't think about.
Somewhere in Nairobi, a worker draws a bounding box around a pedestrian in a photograph. In Manila, another flags a social media post as hate speech. In Kolkata, a third ranks two AI-generated responses to decide which sounds more "helpful." Each task takes seconds. Each pays pennies. And each is essential to the systems that Silicon Valley sells as autonomous intelligence.
Data labeling — the process of annotating images, text, audio, and video so machine learning models can learn from them — is one of the largest and least visible labor markets in the world. It is overwhelmingly performed in the Global South. And the economics are exactly what you'd expect when wealthy nations outsource cognitive work to poorer ones: the same task, the same skill, radically different pay. This is the material reality behind the abstractions of the Training Pipeline — every model that learns to see, read, or reason was taught by human hands.
The Geography of AI Labor
The data labeling industry didn't end up in the Global South by accident. It went where English-speaking workers were available, internet connections were reliable, and wages were low. The pattern follows older outsourcing routes — call centers, back-office processing, garment manufacturing — but with a technological veneer that makes it easier to overlook.
Kenya has become a major hub. Nairobi hosts operations for Sama (formerly Samasource), CloudFactory, and contractors working for Scale AI and Remotasks. Workers label images for computer vision systems, categorize text for natural language processing, and annotate data for companies including OpenAI, Google, and Meta. Kenya's appeal is straightforward: a young, educated, English-speaking population willing to work for wages that are locally competitive but globally negligible.
The Philippines, already the world's call center capital, has added data annotation and content moderation to its outsourcing portfolio. Filipino workers moderate content for Meta, TikTok, and YouTube through contractors like Accenture and Majorel, and increasingly perform data labeling tasks that require cultural fluency in English-language internet culture.
India fields the largest absolute workforce. Companies like iMerit and Karya operate annotation centers across the country. India's data labeling market is projected to grow substantially as demand for Hindi, Tamil, Bengali, and other Indic language training data increases alongside multilingual AI development.
Latin America is the fastest-growing region. Venezuela and Argentina, where economic crises have created large pools of educated workers desperate for dollar-denominated income, have become significant sources of Spanish-language annotation labor. Workers on platforms like Remotasks and Toloka earn $1-3 per hour — low by any global standard, but enough to outpace hyperinflated local currencies.
Madagascar has emerged as one of the newest and lowest-cost hubs. French-language annotation work is routed there, with reported wages sometimes falling below $1 per hour. A 2023 investigation by WIRED documented workers in Antananarivo labeling data for AI systems they would never be able to afford to use (Dzieza, 2023).
The Economics: Wage Arbitrage as Business Model
Strip away the corporate language about "distributed workforces" and "global talent pools" and the data labeling industry's business model is simple: wage arbitrage. Work that would cost $15-25 per hour in San Francisco costs $1-3 per hour in Nairobi, Manila, or Dhaka. The task is identical. The skill required is identical. The pay differs by an order of magnitude.
This is not a side effect. It is the product. Data labeling companies sell AI firms the ability to train models cheaply by exploiting international wage differentials. Scale AI, valued at over $13 billion, built its business on this spread. Sama, which brands itself as an "impact sourcing" company, generates revenue from the gap between what clients pay and what workers earn.
The numbers tell the story. OpenAI's ChatGPT generated billions in revenue in its first full year. The Kenyan workers whose labor was essential to training it — labeling toxic content so the model could learn to avoid it — earned between $1.32 and $2 per hour (Perrigo, 2023). OpenAI paid Sama less than $200,000 for the contract. The model those workers helped refine is now valued in the tens of billions.
That is not a labor market. It is an extraction pipeline. The Economics of Training a Frontier Model rarely accounts for this labor cost, because the industry treats it as negligible — which tells you everything about how the work is valued.
What the Work Actually Involves
Data labeling is routinely described as "unskilled" work. It is not.
Image annotation requires workers to draw precise bounding boxes around objects — pedestrians, vehicles, lane markings, traffic signs — in thousands of photographs. This work trains the computer vision systems used in autonomous vehicles, medical imaging, and surveillance. It demands sustained concentration, spatial reasoning, and consistency across thousands of repetitive decisions.
Text classification asks workers to read passages and make judgments: Is this content toxic? Is this review genuine? Does this response answer the question? These tasks require literacy, cultural knowledge, and interpretive judgment — the ability to distinguish irony from sincerity, critique from abuse, context from content.
RLHF preference ranking — reinforcement learning from human feedback — is the process that makes large language models like ChatGPT feel conversational. Workers read two AI-generated responses and choose which is better. This requires fluency, critical thinking, and the ability to evaluate tone, accuracy, and helpfulness simultaneously. It is, in effect, editorial work. It is compensated as piecework.
Content moderation is the most psychologically demanding category. Workers review flagged content — images and videos depicting violence, sexual abuse, self-harm, hate speech, and child exploitation — and decide whether it violates platform policies. As Sarah Roberts documents in Behind the Screen, this work is traumatic by design (Roberts, 2019). Someone must absorb the worst the internet produces so that platforms remain usable. The scale and severity of this toll is documented in Content Moderation Mental Health.
Every one of these tasks requires genuine cognitive skill. The industry's insistence on calling them "simple" or "routine" serves a specific purpose: it justifies the pay.
The TIME Investigation and Its Aftermath
In January 2023, TIME published an investigation that brought data labeling conditions into mainstream view. Reporter Billy Perrigo documented how Sama workers in Nairobi had been contracted by OpenAI to label toxic content — descriptions of sexual abuse, violence, hate speech — to build the safety filters for ChatGPT. Workers were paid between $1.32 and $2 per hour.
The workers reported psychological trauma. They described being exposed to graphic descriptions of child sexual abuse, bestiality, and extreme violence for hours at a stretch. Counseling was available but insufficient. Some workers reported lasting mental health impacts including anxiety, insomnia, and PTSD-like symptoms. Sama terminated the OpenAI contract early, reportedly due to the severity of the content (Perrigo, 2023).
The investigation revealed something the industry already knew but preferred not to discuss: the safety of the most celebrated AI products is built on the psychological damage of some of the world's lowest-paid workers. The harm doesn't disappear — it is redistributed downward, from wealthy users in the Global North to precarious workers in the Global South.
The Colonial Parallel
Kate Crawford's Atlas of AI (2021) frames the AI industry as an extractive supply chain. Raw materials — lithium from mines in Bolivia, cobalt from the Democratic Republic of Congo, and labor from data centers across the Global South — flow toward a small number of corporations in wealthy nations. The finished products flow back as services sold at enormous margins.
The parallel to colonial resource extraction is not a metaphor. It is a structural description — one that resonates with the broader Digital Divide in AI access and benefit. In colonial economies, raw materials were extracted from the periphery and manufactured into finished goods at the center. The periphery bore the costs — environmental degradation, labor exploitation, political instability — while the center captured the value.
Data labeling follows the same pattern. The raw material is human judgment. It is extracted in Nairobi and Manila, processed in San Francisco and Mountain View, and sold back to the world as artificial intelligence. The workers who provide the essential input have no ownership stake in the output. No equity. No royalties. No credit. In most cases, they don't even know which products their labor supports — non-disclosure agreements ensure that the supply chain remains opaque.
The "Impact Sourcing" Defense
Companies like Sama and CloudFactory market themselves as social enterprises. The argument: they provide employment opportunities in communities where alternatives are scarce. Two dollars an hour is exploitative by American standards, but in Nairobi's Kibera or rural Kenya, it can represent a genuine improvement over available alternatives.
This argument deserves engagement, not dismissal. The workers themselves are not passive victims. Many actively seek data labeling work because it pays better than local alternatives, offers flexible hours, and doesn't require physical labor in difficult conditions. Dismissing their agency or their choices would be its own form of condescension.
But the "impact sourcing" frame obscures a crucial question: why are the alternatives so limited? The same global economic structures that make $2 per hour attractive in Kenya are the ones that make it profitable for Silicon Valley. The data labeling industry doesn't just respond to existing inequality — it is built on it and reinforces it. Workers are paid enough to be retained but not enough to accumulate the resources that would give them bargaining power or alternative options.
The better question is not whether $2 per hour is better than nothing. It is whether the companies generating billions from AI can justify paying the workers who make it possible at subsistence rates. And it is worth noting that some companies are answering that question differently. Not every employer in this space has chosen the extractive model.
What Fair AI Labor Would Look Like
Alternatives exist. They just aren't the default.
Karya, a nonprofit data annotation platform based in India, pays workers approximately 20 times the prevailing local wage for data labeling tasks. More significantly, Karya shares profits from data licensing with its annotators and allows workers to retain rights over the data they produce. The model demonstrates that dignified compensation is technically and economically feasible — it simply requires treating labor as a cost to be fairly borne rather than a margin to be maximized (Karya, 2024).
But Karya is the exception. Making fair labor the norm would require structural changes:
Minimum wage standards for the AI supply chain. If a model is deployed in the United States, the workers who trained it should be covered by labor standards that reflect the value they create, regardless of where they sit. Supply chain labor standards exist in apparel and electronics manufacturing. There is no reason AI should be exempt.
Profit-sharing and equity. Workers whose labor is embedded in AI systems that generate billions in revenue should receive a share of that value. This is not radical — it is the logic behind royalties, residuals, and intellectual property rights applied to a new domain.
Transparency and auditing. AI companies should be required to disclose the labor conditions under which their training data was produced, just as food companies disclose sourcing and apparel companies face supply chain audits. Consumers and enterprise buyers deserve to know whether their AI was trained ethically.
Certification programs. An independent certification for fair AI labor — analogous to Fair Trade certification in agriculture — would give companies a market incentive to improve conditions and give buyers a way to differentiate.
The Redistribution at Stake
The data labeling industry crystallizes the central question of AI redistribution. The technology generates enormous value. That value depends on human labor. And the distribution of returns — billions to shareholders, subsistence wages to workers — is not determined by technology. It is determined by power.
The workers drawing bounding boxes in Nairobi are not peripheral to the AI revolution. They are foundational to it. Every self-driving car, every content recommendation, every chatbot response rests on their labor. The question is whether that labor will be recognized and compensated, or whether AI will reproduce the oldest pattern in global capitalism: wealth flowing upward and outward, costs flowing downward and inward.
The mechanism is visible. The alternatives exist. What remains is the will to choose them.
Projects like Masakhane — building African-language NLP through community-led research — and the collective organizing of the African Content Moderators Union show that workers and researchers in the Global South are not waiting for Silicon Valley to reform itself. They are building alternatives and demanding accountability. The question is whether consumers, policymakers, and AI companies will support those efforts or continue to treat cheap labor as a competitive advantage.
Sources
- Perrigo, Billy. "OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic." TIME, January 18, 2023. TIME
- Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
- Roberts, Sarah T. Behind the Screen: Content Moderation in the Shadows of Social Media. Yale University Press, 2019. Publisher
- Dzieza, Josh. "AI Is a Lot of Work." The Verge, June 20, 2023. The Verge
- Gray, Mary L. and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019. Publisher
- Karya. "Our Model." 2024. Karya