Two Tiers of AI Work, and Africa Sits on the Bottom One
Experts are paid well to teach machines judgment, while African labellers earn a dollar or two an hour to clean up after them.

Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
A South African who recently completed a PhD was offered six hundred rand an hour to help train an AI system to design assessments, teach undergraduates and grade essays, roughly twenty times the national minimum wage. He turned it down. A few thousand kilometres north, Kenyan workers employed through an outsourcing firm were reportedly paid about a dollar fifty an hour to review brutal and traumatising content uploaded to Facebook. In principle, both are doing the same thing: supplying the human judgement that machines cannot yet produce on their own. The difference in what they are paid tells us a great deal about how the AI economy is being built, and where Africa is being placed within it.
CONTEXT AND BACKGROUND
Every model that appears intelligent has been shaped by people. They label images, rank answers, flag violence and correct errors. The organisations that buy AI rarely see this work, partly because it is contracted out through layers of outsourcing firms and gig platforms. A mapping project by African content moderators and researchers found at least one outsourcing centre for moderation or data labelling in thirty-nine of Africa’s fifty-four countries, seventy-eight centres in all. They include a US-headquartered firm operating in South Africa, Nigeria and Kenya.
The conditions at the bottom of this chain are well documented. A preliminary survey of Kenyan data workers found that only two in ten men, and none of the women surveyed, earned enough to cover living costs. Six in ten described their work as mentally exhausting. An independent Oxford audit of Sama, which employs labellers in Kenya and Uganda, cut its fair work score from five to three out of ten. An audit found pay above the minimum wage but below a living wage, no permanent contracts and no recognised worker body.
INSIGHT AND ANALYSIS
What emerges is a two-tier labour market inside AI development. In the upper tier, credentialed experts are paid well to teach models how to reason in specialised fields. In the lower tier, large numbers of young, educated workers perform repetitive and often distressing tasks for a fraction of that pay, on short contracts and with little protection. Both tiers supply judgement. Only one is paid as though judgement were valuable.
Joan Kinyua, founding president of Kenya’s Data Labelers Association, argues that this is no accident. She describes Kenya’s English-speaking, tech-savvy graduates being treated as a source of cheap labour, and contends that the poor conditions are intentional rather than incidental. If she is right, the problem is not that Africa lacks skills. It is that African skills are being bought at the lowest possible price.
This matters for South Africa in two ways. With youth unemployment above forty-seven per cent, as the Rest of World report notes, almost any paid digital work will look attractive, and that weakens workers’ bargaining power. South African companies are also buyers of the AI systems built on this labour, so working conditions are a procurement question as well as a policy one. The country also has a choice about which tier it competes in. The PhD in the opening example shows that South African expertise can attract upper-tier rates. The ambition should be to grow that layer, rather than compete for cheap annotation.
I have previously written about this in the context of training data. I argued that AI vendors should be able to describe what their models were trained on, by category, source and legal basis, as a routine part of contract renewal rather than something discovered through litigation. The same principle should extend to the people who trained the model.
IMPLICATIONS
For policymakers, the choice is about positioning. South Africa’s national AI policy process has already heard calls to make data workers’ right to organise a founding principle. Protections of that kind, combined with investment in specialised skills, would help move African workers from the lower tier towards the upper one, rather than leaving them to compete on price alone.
For boards and executives, the practical step is to treat AI like any other supply chain. Four questions belong in every AI procurement and renewal. Who performed the labelling and moderation behind this model, and through which firms? What contracts and pay did those workers have? What mental health support did those exposed to disturbing content receive? Has any of it been independently audited? A vendor that cannot answer has a supply chain it does not fully understand, and a board that does not ask has accepted that risk without knowing it.
CLOSING TAKEAWAY
AI is often sold as a technology that runs itself. In reality, it runs on human judgement, and that judgement is priced very differently depending on who supplies it and where they live. Africa’s young workers are capable of far more than the cheapest tier of this market, yet that is where most of them are placed today. Whether that changes depends on choices being made now by governments drafting AI policy, companies buying AI systems and educators preparing graduates. The intelligence in these machines is partly African. The rewards should be too.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net




Comments