A doctoral graduate in industrial sociology receives a job offer. Not from a university, but from an AI recruiter. The role: train a machine to design assessments, evaluate student essays, and teach undergraduate courses. The same tasks the candidate spent a decade learning to perform. This is not a hypothetical scenario. It happened to James Maisiri, who completed his Ph.D. at the University of Johannesburg, and his account opens a window onto something that most coverage of AI and employment tends to miss entirely.
The question is not simply whether AI will take jobs. It is whether the people most at risk are being asked to hand over the very thing that makes their work irreplaceable, before any displacement even occurs.
The Knowledge Transfer Nobody Names
There is a well-documented category of AI labor: data labeling, annotation, content moderation. These roles have been outsourced for years to workers in Kenya, Nigeria, and other countries, and they involve relatively mechanical tasks, tagging images, flagging content, transcribing audio.
What Maisiri describes is structurally different. He was not recruited to label data. He was recruited to transfer judgment. Over years of teaching, he had developed the ability to distinguish memorization from genuine understanding, to explain why one essay deserves 75% rather than 60%, to identify which concepts truly matter in an undergraduate course. That accumulated discretion, not just knowledge, was what the AI platform wanted.
Platforms like Outlier, Mercor, and Surge are now actively recruiting doctors, lawyers, engineers, and educators for exactly this reason. White-collar professions have long been considered difficult to automate because they depend on interpreting context rather than applying fixed rules. That assumption is now being tested, and the testing is being done by the professionals themselves.
47.4% Unemployment and a 600-Rand Offer
The economic dimension of this story is impossible to separate from the ethical one. South Africa’s youth unemployment rate stood at 47.4% in the second quarter of 2026. The national minimum wage is 30.23 rand per hour, roughly two US dollars. The AI training role Maisiri was offered paid 600 rand per hour, equivalent to about 37 US dollars. In that context, the decision to participate or refuse is not a philosophical exercise. It is a livelihood calculation.
This is precisely the dynamic that makes Africa particularly exposed to what might be called knowledge extraction at scale. The continent has one of the lowest AI adoption rates globally, with most countries below the global average of around 18% among the working-age population. South Africa sits slightly above that, at 23%. Yet Africa also has some of the youngest populations in the world, a growing pool of highly educated professionals, and an economic environment defined by high unemployment and modest growth. For AI companies, that combination creates a structural incentive: expert knowledge available at costs significantly lower than in Western markets.
Meanwhile, at the other end of the pay scale, waste sorters and welders in India are being paid 250 rupees per hour, about 2.60 US dollars, to wear cameras and capture their daily work routines. That footage trains humanoid robots designed to perform the same jobs. The wage gap between Maisiri and those workers is enormous. The underlying transaction is identical: human expertise, embodied or intellectual, converted into machine capability.
What Judgment Actually Means, and Why It Cannot Be Separated from the Person
Here is what most discussions of AI and labor fail to address directly. Judgment is not a skill in the conventional sense. It is not a technique that can be documented, extracted, and transferred cleanly. It is the product of accumulated experience, ethical formation, and contextual sensitivity. When a teacher decides that one student’s essay demonstrates genuine understanding while another’s does not, that decision draws on everything the teacher has seen, read, and taught over years. It is inseparable from who they are.
When AI platforms recruit professionals to train their systems, they are not just acquiring data points. They are attempting to encode that kind of situated, personal knowledge into a model that will then exercise it without the person. Maisiri’s account raises a question that deserves to be taken seriously: what kind of judgment are these systems actually learning? Is it fair judgment? Is it ethical judgment? Does it recognize context in the way a human practitioner would?
These are not rhetorical questions. They are design questions, and right now, the people best positioned to answer them are being hired to provide inputs, not to shape outcomes.
Maisiri ultimately walked away from the process, though he acknowledges he cannot fully explain why. He continues to look for an academic position. His story does not resolve the tension it describes. It simply makes it visible.
In Short
AI systems are increasingly being trained not on raw data but on professional judgment: the kind that takes years to develop and defines entire careers. Highly educated workers in high-unemployment economies face a genuine dilemma, because the financial incentive to participate is real, and so is the risk of accelerating their own displacement. What is being transferred in these transactions is not just knowledge. It is the capacity to decide, and once transferred, it does not come back.
Based on reporting from Rest of World.