Those small-town youths labeling AI large models

By: Sleepy.md

Datong in Shanxi—once a city propped up by coal that powered half the country—has now shaken off the coal dust covering its body, picked up a sharper pickaxe, and slammed it hard into another faceless mine.

Inside the office buildings in the Jīn Mào International Center in Pingcheng District, there are no more hoisting shafts, and no more coal-hauling trucks. In their place are thousands of tightly arranged computer workstations. The Shanghai Runxun Yunzton Voice Valley big data intelligent services base takes up several full floors. Thousands of young employees wearing headsets are staring at their screens, clicking, dragging, and selecting boxes.

According to official data, as of November 2025, Datong has put into operation 745k servers, brought in 69 call center data annotation companies, and created employment close to home for more than 30k people. Output value is 750 million yuan. In this digital pit full of numbers, 94% of those working in the industry have local household registration.

It’s not only Datong. Among the first batch of data annotation bases confirmed by the National Data Bureau, places like Yonghe County in Shanxi, Bijie in Guizhou, and Mengzi in Yunnan are clearly on the list. At the Yonghe County data annotation base, 80% of employees are women. Most of them are rural stay-at-home moms, or young people who returned to their hometowns because they couldn’t find suitable work.

A hundred years ago, in a textile mill in Manchester, England, it was packed with farmers who had lost their land. And today, in front of the computer screens in these remote county towns, sit young people who can’t find a place for themselves in the real economy.

They are doing a task that feels intensely futuristic, yet is extremely primitive and paid per piece—producing the data feed that AI giants in Beijing, Shenzhen, and Silicon Valley need to build large models.

Nobody thinks there’s anything wrong with that.

A New Assembly Line on the Loess Plateau

The essence of data annotation is to teach machines to recognize the world.

Autonomous driving needs to recognize traffic lights and pedestrians; large models need to figure out what is a cat and what is a dog. Machines themselves have no common sense. Humans must first draw a box on the image, telling it, “This is a pedestrian,” and only then, after swallowing tens of millions of images, can it learn to recognize on its own.

This job doesn’t require a high level of education—only patience, and a single finger that can keep clicking nonstop.

In the golden era of 2017, a simple 2D box could cost more than a dime per unit—so much that some companies even offered as much as five cents. Labelers with fast hands could work more than ten hours a day and earn five or six hundred yuan. In a county town, that’s absolutely a high-paying, respectable job.

But as large models evolved, the brutal side of this assembly line began to show.

By 2023, the unit price for simple image annotation had already been crushed to 3–4 cents. That’s a drop of more than 90%. Even for the more difficult 3D point cloud images—those made of dense points that require magnification countless times to see the edges clearly—annotators still had to draw a 3D box in three-dimensional space that includes length, width, height, and the yaw/pitch/roll angles, tightly enclosing the vehicle or pedestrian. And even this complex 3D box was only worth five cents.

The direct consequence of the unit price collapsing is a massive increase in labor intensity. To tightly hold onto the monthly base salary of two or three thousand yuan, annotators have to constantly, nonstop, improve their hand speed.

This is not some easy white-collar job. In many annotation bases, management is strict to the point of suffocating. During work hours, you’re not allowed to answer phone calls; your phone must be locked inside a storage locker. The system precisely records each employee’s mouse movements and dwell time. If you pause for more than three minutes, warnings from the backend will come at you like a whip.

What’s even more maddening is the tolerance rate. The passing threshold in the industry is usually above 95%. Some companies even require 98%–99%. That means if you draw 100 boxes and get 2 wrong, the entire image will be sent back for rework.

Video data is continuous frame by frame. When a vehicle changes lanes, it can be occluded, and annotators must use inference to find each one; in 3D point cloud images, any object with more than 10 points must be boxed. For a complex parking-lot project, if the lines are drawn too long or you miss labels, quality inspection will always pick up the problems. Getting a picture sent back for rework four or five times is routine. In the end, after spending about an hour, what you take home is only a few cents.

A data labeler in Hunan posted her settlement slip on a social platform. After a day’s work, she drew more than 700 boxes, the unit price was 4 cents, and her total income was 30.2 yuan.

This is an extremely split picture.

On one side are the flashy technology executives at product launches, talking about how AGI will liberate humanity. On the other side are young people in county towns on the Loess Plateau and in the southwestern mountains, staring at screens for eight to ten hours a day—mechanically drawing boxes, thousands, tens of thousands, and even while dreaming at night, their fingers still drawing lane lines in midair.

Someone once said that the outward appearance of artificial intelligence is like a luxury car roaring by—but if you open the door, you’ll find that inside, a hundred people are riding bicycles, clenching their teeth and pedaling as hard as they can.

Nobody thinks there’s anything wrong with that.

A Piecework Job Teaching Machines “How to Love”

Once the bottleneck of image recognition is broken through, large models move into even deeper evolution: they need to learn to think, talk, and even to show “empathy” the way humans do.

This creates the most core—and most expensive—stage in large-model training: RLHF (reinforcement learning from human feedback).

In simple terms, it means letting real people score AI-generated answers, telling it which response is better and more aligned with human values and emotional preferences.

ChatGPT looks “human” because behind the scenes there are countless RLHF annotators teaching it.

On crowdsourcing platforms, these annotation tasks are often priced transparently: per item, 3 to 7 yuan. Annotators have to give AI responses extremely subjective emotional scores—judging whether the response is “warm,” whether it is “empathetic,” and whether it “takes the user’s emotions into account.”

A bottom-tier worker with a two- or three-thousand-yuan monthly wage, struggling to survive amid the mud of reality—who has no time even to care about their own emotions—must, in the system, serve as an AI’s emotional mentor and a values judge.

They have to forcibly tear apart these extremely complex and subtle human emotions like warmth and empathy, then quantify them into cold scores from 1 to 5. If their scoring doesn’t match the system’s pre-set standard answers, they’re judged as not meeting the accuracy requirement, and their originally meager piecework pay will be deducted.

This is a form of cognitive extraction. Humans’ complex and delicate emotions, morality, and compassion are being dragged into the algorithm’s funnel by force. In the cold world of quantification and standardized scales, they are squeezed dry of their last bit of warmth. When you marvel that the cyber beast on the screen has learned to write poetry and compose music, to show concern and ask after others’ wellbeing, and even to put on a cloak of melancholy—outside the screen, that group of originally vibrant humans is, day after day, reduced by mechanical judgment into emotionless scoring machines.

This is the most hidden side of the entire industrial chain—something that never appears in any financing news or technical white papers.

Nobody thinks there’s anything wrong with that.

985 Master’s Degrees and Town Youth

The piecework of pulling boxes at the bottom is being crushed under the treads of AI. This cyber assembly line is beginning to spread upward, starting to devour even higher-level mental labor.

The appetite of large models has changed. It no longer satisfies itself with chewing up simple common sense—it needs to swallow human professional knowledge and higher-level logic.

On major job platforms, a special kind of part-time work is now flashing frequently, such as “large-model logic reasoning annotation” and “AI humanities training instructor.” The threshold for this part-time work is very high, often requiring “985/211 master’s degree or above,” and it covers professional fields like law, medicine, philosophy, and literature.

Many top-university graduate students are attracted and flow into these outsourcing groups at big tech companies. But soon they realize this isn’t really some easy mental exercise—it’s a spiritual torment.

Before taking official orders, they must read documents dozens of pages long detailing scoring dimensions and evaluation standards, and do two to three rounds of trial annotations. If they pass, then during official annotation, if their accuracy is below the average level, they lose the qualification and are kicked out of the group chat.

Most suffocating of all is that these standards are basically not fixed. When faced with similar problems and answers, scoring them using the same way of thinking can still yield totally opposite results. It’s like taking a test that’s never-ending, with no real standard answers. You can’t improve accuracy through personal effort or study—you can only endlessly circle in place, consuming brainpower and physical stamina.

This is the new exploitation in the era of large models—class compression.

Knowledge, which was once seen as the golden ladder for breaking barriers and climbing upward, has now fallen to become digital fodder offered to algorithms—fodder that is even more complex to chew. Under the absolute power of algorithms and systems, both the 985 master’s students in ivory towers and the town youth on the Loess Plateau have arrived at the most bizarre convergence of different roads to the same destination.

Together, they tumble into this bottomless cyber mine, stripped of their halo, flattened into sameness, and turned into cheap gears on the conveyor—gears that can be replaced at any time.

It’s the same overseas. In 2024, Apple cut an AI voice annotation team of 121 people in Santiago outright. These employees were responsible for improving Siri’s multilingual processing capabilities. They once thought they were standing at the edge of a big company’s core business—but in an instant, they fell into the abyss of unemployment.

To tech giants, whether it’s the box-drawing auntie in a county town or the logic training instructor who graduated from a top school, in essence they’re all “consumables” that can be replaced at any time.

Nobody thinks there’s anything wrong with that.

A Trillion-Trillion Babel, Built with Blood Sweat Worth a Few Cents

According to data released by China’s Academy of Information and Communications Technology, in 2023 China’s data annotation market size reached 6.08 billion yuan. In 2025 it is expected to reach 20 to 30 billion yuan. And according to forecasts, by 2030, global sales revenue for data annotation and services will surge to 117.1 billion yuan.

Behind these numbers is a frenzy of valuations among tech giants like OpenAI, Microsoft, and ByteDance—often worth trillions or tens of trillions of dollars.

But none of this huge wealth has flowed to the people who truly “feed” AI.

China’s data annotation industry has a typical inverted pyramid outsourcing structure. At the very top are the tech giants who tightly control the core algorithms. The second layer consists of large data service providers. The third layer is made up of data annotation bases across the country and small-to-medium outsourcing companies. At the very bottom are the low-end “dirt leg” annotators who get paid by the piece.

For each layer of outsourcing, they take a big cut of the profits. When the big company sets the unit price at five cents, after being stripped layer by layer, what finally lands in the hands of a county-town annotator may be less than even five cents.

Greek former finance minister Yanis Varoufakis put forward a highly penetrating viewpoint in his book Technology Feudalism: today’s tech giants are no longer capitalists in the traditional sense—they are “Cloudalists.”

They don’t own factories and machines; they own algorithms, platforms, and computing power—these are digital territories in the cyber era. In this new feudal system, users are not consumers but digital serfs. Every like, comment, and browse you do on social media is freely supplying data to the Cloudalists.

And the data annotators distributed in lower-tier markets are the system’s lowest-level digital slaves. They don’t just produce data—they also have to clean, categorize, and score massive amounts of raw data, turning it into high-quality feed that large models can digest.

This is a hidden operation to claim territory in cognition. Just like the enclosure movement in 19th-century England drove farmers into textile mills, today’s AI wave drives those young people who can’t find a place in the real economy into the screens.

AI hasn’t erased the class divide. Instead, it has built a “data and sweat conveyor belt” stretching from county towns in central and western China all the way to the headquarters of big tech giants in Beijing and beyond—Beijing, Shanghai, Guangzhou, Shenzhen, and their equivalents. The narrative of technological revolution is always grand and dazzling, but its underlying reality is always large-scale consumption of cheap labor.

Nobody thinks there’s anything wrong with that.

No More a Tomorrow for Humans

The most brutal ending is coming soon—and it’s coming faster and faster.

As large models’ capabilities leap, the annotation tasks that once required humans working day and night are being taken over by AI itself.

In April 2023, Li Xiang, founder of Ideal Automobile, revealed data in a forum. Previously, Ideal had to do about 10 million frames of automated driving image annotation per year; the outsourcing cost was close to 8B yuan. But when they used large models for automated annotation, what once took a year could basically be completed in just three hours.

Efficiency is 1,000 times that of humans—and it was already in 2023. In the just-passed March, Ideal also released a new-generation MindVLA-o1 automated annotation engine.

In the industry, there’s a joking line that’s extremely true: “As much intelligence as there is, there is as much manual labor.” But now, in big companies’ spending on data annotation outsourcing, there has already been a cliff-like drop of 40%–50%.

Those town youth who have spent countless day and nights in front of computers, burning their eyes red, personally fed a giant beast. And now, this beast has turned its head, smashing their livelihoods.

As night falls, office buildings in Pingcheng District of Datong remain as bleakly white as day. Young people on shift handoff silently exchange their exhausted bodies in the elevator lobby. In this folded space tightly imprisoned by countless polygon boxes, no one cares what kind of epic leap the Transformer architecture across the ocean has achieved, and no one can make sense of the roaring behind hundreds of billions of parameters of compute.

Their attention is only welded to that red-green progress bar in the backend that represents the “passing line,” calculating whether those few points and few cents of piecework numbers can be pieced together into a decent life by the end of the month.

On one side are the Nasdaq bell-ringing moments and a flood of technology media coverage—giants toasting the arrival of AGI. On the other side are these digital serfs who have fed AI with their own flesh and blood one mouthful at a time. They can only tremble and wait in sour, painful dreams for the giant beast they raised with their own hands to, on some seemingly ordinary morning, casually kick away their meal ticket.

Nobody thinks there’s anything wrong with that.