From Kenya’s data annotators to African AI startups, the continent contributes essential work to global models. But who captures the skills, ownership and career progression?
The key points of this analysis, checked by the editors.
AI labour is a valid entry point only if worker rights and career progression exist.
Africa needs ownership, technical leadership and a larger share of the value chain.
A model responds in polished English. A driverless-car system identifies a traffic light. A chatbot refuses a dangerous request. Behind each demonstration sits a less visible supply chain: people collecting examples, labelling images, assessing responses and filtering disturbing content. Some of that work is done in African cities and refugee communities. The industry's favourite word for the opportunity is inclusion. Workers have a harder question: inclusion at which level of the value chain?
The work is real, and so are the gaps
Kenya offers a useful window into the market. An International Labour Organization study documents how online platforms have opened income opportunities while leaving workers exposed to uneven pay, weak legal protection and unequal conditions. In its surveyed microtask segment, respondents reported work on artificial intelligence, machine learning, data collection and verification. The study also found disadvantages affecting women workers, including lower earnings in many areas and reduced access to social protection. Read the ILO report.
This is not an argument that data annotation is worthless. Annotation can be an entry point into digital employment, an important source of cash, and a way to learn quality standards. But there is a gulf between becoming proficient at a repetitive task and obtaining the power to design the system that assigns, measures and purchases that task.
Ask who owns the contract, who sets the quality threshold, who sees the model's commercial performance, who can negotiate the next rate, and whether the worker is credited for expertise. Those questions reveal whether a job is the first rung of a ladder or a treadmill.
The part of an AI product that rarely appears in the launch video
Consider an image-labelling assignment. A worker may be paid to identify road markings for autonomous-vehicle training. They might spend hours annotating thousands of frames, resolving ambiguities the software cannot yet handle. This effort improves a product that may later be sold into far wealthier markets. The worker's payment compensates labour; it does not automatically confer intellectual-property rights, equity or participation in the product's upside.
The asymmetry is structural, not proof that every outsourcing company behaves badly. A supplier operating under a transparent contract, providing counselling for harmful-content exposure and promoting experienced reviewers can provide valuable employment. Conversely, a gig platform can advertise « global opportunity » while hiding rejection criteria, delaying payments and allowing unilateral account suspension.
The ILO has argued that extending labour and social protections to platform workers is essential for a healthy digital economy. Workers classified as independent contractors can otherwise fall between existing labour protections. ILO's Kenya policy discussion.
Training programmes should be judged by where graduates land
African governments and private organisations frequently announce cohorts trained in AI. A headcount is easy to report; a durable career outcome is harder. Six months after graduation, how many participants have contracts? How many receive a predictable income? How many have moved from annotation to quality assurance, data engineering, model evaluation, security or product management?
That transition matters because even annotation work is exposed to automation. AI can generate labels for other AI systems, with humans reserved for difficult cases. A workforce trained only for the cheapest task risks being squeezed precisely when its productivity rises.
A stronger programme would teach dataset versioning, Python fundamentals, SQL, quality sampling, privacy, red-teaming and evaluation design. Workers would learn to audit outputs and manage projects, not simply click boxes faster. The curriculum should include contracts, client negotiation and the difference between a gross task rate and income after internet, hardware and unpaid waiting time.
Ownership cannot be outsourced forever
The higher-value segments of the chain involve deciding which problems to solve, securing reliable local data, designing products, operating infrastructure and selling services. That means African universities, businesses and investors need more than a promise to supply training labour.
There are practical starting points. Healthcare providers can co-design privacy-aware local tools instead of sharing sensitive records with unvetted vendors. Agricultural cooperatives can build datasets with documented consent and benefit-sharing. African-language technologies can compensate language experts as contributors with specialist knowledge, not interchangeable anonymous clicks.
For startups in Accra, Nairobi, Kigali and Douala, a viable AI strategy need not mean training a frontier model from scratch. It can mean solving a specific regional problem, owning the customer relationship, building trusted evaluation datasets and retaining the expertise to switch infrastructure providers when needed.
Five questions for every new « AI jobs » announcement
The most useful public scoreboard would show pay transparency, verified contracts, worker protections, skills progression and the number of local managers or product owners promoted from entry-level teams. It should also report whether workers can dispute automated rejections and whether psychological support exists for moderation roles.
There is nothing wrong with starting in the global AI supply chain. The problem arises when the starting point is marketed as the destination.
Africa is already helping AI systems learn. The next contest is whether enough Africans get to decide what those systems are built for — and share in the value they create.
Sources and method
Facts that may change over time were checked against the sources below. Observations and interpretations are presented as such in the article.
