Experts Are Being Paid to Train the AI Systems That Could Replace Them
After completing a PhD on AI’s reshaping of agricultural work in South Africa, a researcher was offered her first job. Not an academic role. A recruiter invited her to train an AI system to design assessments, teach undergraduates, and mark essays — everything she had spent a decade learning to do herself. The pay was 600 rand per hour, approximately $37.
She turned it down. Most people facing the same offer are not in a position to.
The Economics of Replacement
South Africa’s national minimum wage is 30.23 rand per hour — $2 at current exchange rates. Youth unemployment stood at 47.4% in the second quarter of 2026. At $37 an hour, the offer was roughly 20 times the legal floor in a country with near-half youth unemployment. The gap between principle and economic reality is not abstract.
In India, the arithmetic looks different but the structure is identical. Waste sorters and welders are strapping phones or cameras to their heads for 250 rupees per hour — $2.60 — to capture the physical mechanics of their own work. That data trains the humanoid robots that could eventually perform the same jobs without human operators.
At both ends of the pay scale, the activity is the same: transferring human knowledge and judgment to a machine, with the human as the source of value being extracted.
Africa as the Extraction Zone
Africa’s position in this dynamic is structurally specific. The continent has among the youngest and fastest-growing educated workforces in the world, with widespread unemployment and low prevailing wages. AI adoption sits around 18% globally among working-age populations; South Africa is at 23%, slightly above the average but still well below the United States, Europe, or East Asia.
For AI companies, the combination is commercially attractive: deep domain expertise available at a fraction of the cost charged by Western professionals doing the same work. Lawyers, academics, translators, and engineers in African markets represent exactly the kind of high-quality labeled training data that can be acquired at a discount relative to the sophistication it encodes.
The extraction does not require bad faith. AI companies are paying above local market rates. Workers are making choices that improve their immediate economic position. The issue is structural: the knowledge being transferred has a recipient (the model), a buyer (the lab), and a seller (the worker), but the worker is the only party whose future employment the asset may displace.
What the Pattern Looks Like at Scale
Skilled writers across multiple markets are being paid to “humanize” AI-generated text — editing outputs so they read as if produced by a person. Translators are post-editing machine translations. Radiologists are labeling scans to train diagnostic models. Software engineers are reviewing AI-generated code to produce clean training signal.
Each of these roles requires the specific expertise being transferred. The labor is not unskilled. What it lacks is leverage: no union, no recognized trade, no regulatory category, and no obvious substitute income if the work succeeds at its stated purpose and the market for the human skill contracts.
What This Data Point Represents
The South African researcher’s account is notable not for its rarity but because she refused and published the reasoning. The overwhelming majority of workers in this pipeline accept the work and do not publish. The data point visible in the Rest of World article is one she chose to surface. The size of the pool of people who chose differently, said nothing, and transferred their expertise is unknown.
The $37-an-hour offer for academic-equivalent AI training work, set against $2-an-hour minimum wages and 47% youth unemployment, is not a manipulative edge case. It is the system working as designed.