By Sleepy.md
In Datong, a northern Chinese city once defined by coal, a new kind of extraction is under way. Gone are the shafts and coal trucks. In their place, inside office towers, rows of computer terminals stretch across entire floors. Thousands of young workers wearing headsets sit clicking, dragging, and boxing images on their screens. The city, long reliant on fossil fuels, is now digging down into a different pit: data.
By late 2025, Datong had attracted 69 data labeling companies employing more than 30,000 people and generating 750 million yuan ($100 million) in output. Similar hubs have sprung up across inland China, from Shanxi in the north to Guizhou and Yunnan in the southwest, often staffed by locals — many of them women or returning migrant workers.
Their task is simple but essential: teaching machines how to see and think. For an autonomous vehicle to recognize a pedestrian, a human must first draw a box around an image and label it; large language models must distinguish cats from dogs. Feeding millions of labeled examples into algorithms requires patience rather than formal education — and a finger capable of relentless clicking.
To maintain a monthly salary of 3,000 yuan, workers must achieve grueling speeds. This is not white-collar work. Management is stifling: phones are locked away, and systems track mouse movements in real-time. A three-minute pause triggers a warning. Accuracy requirements are often set at 99%. A single mistake in a batch of 100 boxes can result in the entire task being sent back for unpaid reworking.
One worker in Hunan recently shared her earnings: for a full day’s work labeling 700 frames at 0.04 yuan each, she earned a total of 30.2 yuan ($4.15). While tech moguls discuss how Artificial General Intelligence (AGI) will liberate humanity, these workers spend 10 hours a day mechanically drawing lines until they see them in their sleep.
AI may look like a high-speed car, but inside, hundreds of people are frantically pedalling.
A sweatshop for graduates and farmers
As image recognition matures, the industry has moved toward Reinforcement Learning from Human Feedback (RLHF). This requires humans to rank AI responses based on empathy and warmth to make bots like ChatGPT sound more human.
On crowdsourcing platforms, these tasks pay between 3 yuan and 7 yuan. It is a profound irony: workers struggling in the mud of reality, often unable to afford their own emotional health, are tasked with being ethics tutors for AI. They must quantify complex human emotions into cold scores of 1 to 5. If their subjective empathy score deviates from the system’s standard, their pay is docked.
This cyber-assembly line is now moving up the value chain. Large Language Models (LLMs) now require “high-order logic” training, attracting graduates from China’s elite universities. These specialists—holding masters degrees in law or philosophy—find themselves subjected to the same algorithmic tyranny as the rural workers. They navigate 50-page grading manuals only to be kicked from the group if their accuracy fluctuates.
The industry is expanding rapidly. China’s data labeling market was worth an estimated 6.08 billion yuan in 2023 and was forecast to quadruple to around 30 billion yuan in 2025. Worldwide, the sector is projected to exceed 100 billion yuan by 2030, underpinning the soaring valuations of companies such as OpenAI, Microsoft and ByteDance.
Technofeudalism rules
Yet the gains are unevenly distributed. The sector operates through layers of outsourcing: from tech giants to large contractors, then to regional firms and local centres. At each stage, margins are stripped away, leaving frontline workers with only a fraction of the original contract value.
Yanis Varoufakis, an economist and former finance minister of Greece, has described this as technofeudalism, where tech giants are cloudalists, the ruling class who own the digital territory, while labelers are digital serfs providing the labor.
The final irony is that this labor may itself be temporary as AI starts to label itself. Li Xiang, founder of Li Auto, noted that automated labeling has reduced tasks that took a year to just three hours. As these digital serfs finish feeding the beast, the beast is turning around to eat their jobs.
In Datong’s office towers, the lights remain on late into the night. Workers change shifts in silence, eyes fixed on performance dashboards that determine whether they will earn enough for the month. Elsewhere, executives celebrate the arrival of ever more powerful AI. What is rarely acknowledged is the vast, largely invisible workforce whose contribution rarely appears in earnings reports or product launches, but without whom the system could not function.
Source:
Beating
