The New AI Data Job Is Not Labelling Images. It Is Teaching a Model How You Decide
A South African sociologist writes about being recruited to train an AI to design assessments and mark essays, for 600 rand an hour in a country where the minimum wage is 30. The interview was conducted by an AI.
James Maisiri finished a PhD in industrial sociology at the University of Johannesburg this year, on how technologies like AI are reshaping agricultural work in South Africa. His first job offer did not come from a university. A recruiter invited him to train an AI system to design assessments, teach undergraduates, and mark essays. He wrote about it for Rest of World last week.
The money explains why he took the interview. The role paid 600 rand an hour, about $37, in a country where the national minimum wage is 30.23 rand, about $2, and where youth unemployment sat at 47.4 percent in the second quarter of 2026. As he puts it, when earning a living collides with your principles about helping build a technology that may compete with you, the choice is not obvious.
What makes his account worth reading is the distinction he draws. The data work outsourced to Kenya, Nigeria, and elsewhere over the past decade was labelling: this image contains a cat, this comment is abusive. What he was recruited for is different. Over a decade he learned which concepts actually matter in a course, how to tell memorisation apart from understanding, and why an essay earns 75 percent rather than 60. The model would not just learn what he knew. It would learn how he decided.
What is behind this
That is the part that changed. Professional judgment, the discretion a doctor, lawyer, or teacher exercises, was long considered hard to automate because it depends on reading context rather than applying rules. Platforms like Outlier, Mercor, and Surge now actively recruit exactly those professionals, because judgment turns out to be teachable if you can pay enough qualified people to demonstrate it.
Maisiri sets his situation next to workers in India strapping cameras to their heads for 250 rupees, about $2.60, an hour so humanoid robots can learn welding and waste sorting. Different pay scale, same transaction. Africa is exposed here in a specific way: among the youngest populations in the world, a growing pool of highly educated professionals, high unemployment, and AI adoption below the global average of roughly 18 percent. That combination makes expert knowledge cheap to buy.
His own recruitment had a fitting detail. No human was involved. An AI interviewed him for 45 minutes, emailed him feedback on his strengths and weaknesses, and suggested he retake part of the assessment. He did not, and stopped pursuing the job. He is honest that he still does not fully understand his own discomfort.
What this means for you: If your work depends on judgment you built over years, it is worth knowing that this judgment is now a purchasable input, and that the offers will be aimed at the people who need the money most. If you use AI professionally, the useful question is the one Maisiri asks: whose judgment is in this model, under what conditions was it collected, and were they in a position to say no.
Sources
Source: https://restofworld.org/2026/ai-training-jobs-expert-replacement/
World Labs Shows Atlas, a Model That Builds Walkable 3D Scenes From a Handful of Photos
Fei-Fei Li's company unveiled a model that generates, reconstructs, and simulates 3D space rather than flat video. It outputs point clouds, runs a minute of 1440p footage, and can produce training data for robots.
