There is something almost dizzying about the scene. An architect trained for years finishes her day, opens an online platform and explains to a machine, step by step, how to prepare a building proposal or conduct a profitability analysis. This is not a classroom exercise. She draws on her real work, her professional habits, and the part of expertise that is rarely learned from a manual.

The architect is called Cuicui and lives in Shenzhen. Over five years, her income has been cut in half amid a decline in public infrastructure spending. In the evenings, she therefore works for TalentsAI to pay her mortgage and her child's extracurricular activities. The paradox is clear: in the short term, she is monetizing knowledge that could, in the longer term, help make her own profession less scarce.

To stay solvent today, some professionals are teaching machines what could make their own work less indispensable tomorrow.

This is no longer annotation: it is professional knowledge transfer

The word “annotation” still conjures up small repetitive tasks: classifying a photo, checking a sentence, clicking on a category. Here, the scale is different. Platforms are looking for people who can convey professional reasoning, methods, trade-offs, and sometimes even a way of recognizing a good answer without being able to explain it through a simple rule.

ByteDance launched Xpert in 2025 and says it has recruited more than 50,000 experts, from writers to therapists. Alibaba, through Siriser, recruits teachers, mechanical engineers and composers, among others. TalentsAI and MeetChances also target journalists, tax specialists and headhunters. In the United States, Mercor, Surge AI and Handshake are already recruiting lawyers, scientists and film professionals to train models on white-collar tasks.

The shift matters: the raw material is no longer just data, but human judgment. In these assignments, what companies want is the way an expert moves from an incomplete problem to an actionable decision. In June, China's National Data Administration also called for the development of high-quality datasets and encouraged university graduates to consider annotation as flexible work.

China's AI training data market is expected to reach $1.1 billion in 2026, 25% more than in 2025. At the same time, daily token consumption in China rose from around 100 billion in early 2024 to 500 trillion by mid-2026. These figures do not mean experts are going to disappear. They do show how much the ecosystem needs a more sophisticated kind of fuel: examples of reasoning that models can use.

Why agree to train a potential competitor?

The simplest answer is also the least comfortable: because a moral principle rarely carries as much weight as a mortgage payment coming due. In China, workers interviewed describe this work as demanding, precarious and less well paid than in the United States. Some assignments pay between 100 and 500 yuan, or roughly $15 to $74, while requiring several hours. Tasks can be rejected by quality teams and go unpaid.

The economic context therefore plays a central role. Youth unemployment in China is reported at 17.9%. MeetChances even used a very direct advertising pitch on RedNote: after being laid off in middle age, becoming an AI trainer can be an option. The promise is not a prestigious career. It is income available now.

The same conflict appears elsewhere. In South Africa, James Maisiri, who holds a doctorate in industrial sociology, refused an assignment in which he was asked to transfer to an AI what he had learned over ten years about assessment design, teaching and grading. The role paid 600 rand an hour, about $37, far above the national minimum wage of 30.23 rand. Yet he concluded that the compensation did not make up for the underlying issue: helping to build a tool capable of competing with the profession he had spent years mastering.

Not everyone can afford to refuse. In South Africa, youth unemployment reached 47.4% in the second quarter of 2026. That is where the dilemma becomes less philosophical. The more the labor market deteriorates, the harder it becomes to hold the line between “I protect my expertise” and “I sell what I can sell.”

Better-paid expertise does not remove the paradox

In the United States, the model can be far more lucrative. Mercor, valued at $10 billion, has more than 30,000 white-collar contractors working for labs such as OpenAI or Anthropic. Lawyers and doctors there reportedly earn an average of $85 to $200 an hour, while some specialties, such as psychiatry or venture capital, can reach $350 to $1,000 an hour.

But a high rate does not change the nature of the exchange. The expert is selling less an hour of production than a piece of their comparative advantage. That is precisely why this new market deserves to be viewed as something more than a simple “AI side job.”

What China signals — and does not signal — for Europe

We need to be precise: none of the cases documented here is enough to claim that this phenomenon is already established in France or Europe. China, the United States, South Africa and India do show, however, that the mechanism is not unique to a single economic system. It appears both in markets where expertise is highly paid and in countries where necessity pushes people to accept precarious assignments.

What can be transposed to Europe is therefore not necessarily the level of pay, the legal status, or even the exact form of the platforms. It is the mechanism: when a sector hires less, an AI company seeks a rare skill, and a professional has that skill but lacks opportunities, a market can emerge to turn tacit knowledge into training material.

This possibility deserves even more attention because entry-level opportunities for some white-collar workers — young lawyers, financial analysts and researchers — have reportedly fallen by 50% since 2019, according to Datasumi. It would be excessive to conclude that AI is the sole cause or that Europe will mechanically follow the same path. But the sequence is coherent enough to raise a serious question: what happens when the first jobs that once allowed people to gain experience become scarcer, while machines are learning precisely how to perform some of those tasks?

Human knowledge may be worth more as raw material than as a job

To me, this is the most troubling point. For a long time, we assumed complex professions would be protected because they rely on years of experience, context, exceptions and judgment. Yet it is precisely this complexity that is becoming valuable to AI companies.

Human know-how has therefore not become useless. Almost the opposite: it is valuable enough that experts are paid to capture it. The question is who will retain the value once that knowledge has been transformed into reproducible software capability.

  • For the professional, the assignment provides immediate income.
  • For the platform, it turns rare expertise into examples that can be reused at scale.
  • For the client company, it may reduce the future need to involve an expert in every similar task.

This does not mean every lawyer, architect or engineer is programming their own dismissal. Nothing in the documented cases supports such a mechanical conclusion. Yet the dynamic deserves better than the cliché of the “robot stealing jobs.”

The real shift is more discreet: professionals under pressure can become temporary suppliers of the knowledge that will automate part of their own value chain. China mainly shows us what this market becomes when economic pressure is intense. For Europe, the warning is not a prophecy. It is a question to settle before necessity decides on behalf of professionals: what is expertise worth when it is sold once, but can then be reproduced almost endlessly?