Esos jóvenes de pueblos pequeños que etiquetan grandes modelos de IA

Autor: Sleepy.md

Datong, in Shanxi, a city that was once propped up by coal and accounted for half of the region’s fortunes, has now shaken off the coal dust all over its body, picked up a sharper chisel, and slammed it heavily down onto another invisible mine.

Inside the office buildings in the Jinguo International Center in Pingcheng District, there are no longer elevator shafts, and no more coal-transport trucks. In their place are thousands upon thousands of computer workstations arranged tightly in rows. The Shanghai Runxun Cloud Voice Valley Big Data Smart Service Base occupies several entire floors. Thousands of young employees wearing headsets are staring at their screens, clicking, dragging, and box-selecting.

According to official data, as of November 2025, the city of Datong has put 745,000 servers into operation, and has brought in 69 call-labeling data enterprises, enabling more than 30,000 people to find nearby employment, with an output value of 750 million yuan. In this data-mine pit of numbers, 94% of practitioners are local residents with local household registration.

Not only Datong. Among the first batch of data-labeling bases determined by the National Data Bureau, counties in the central and western regions such as Yonghe County in Shanxi, Bijie in Guizhou, and Mengzi in Yunnan are clearly on the list. In the data-labeling base in Yonghe County, 80% are female employees. Most of them are rural moms, or young people who returned to their hometown because they couldn’t find suitable work.

A hundred years ago, in Manchester’s textile factories in the UK, farmers who had lost their land were packed in. But today, in front of computer screens in these remote county towns, young people occupy the seats who can’t find a place in the real economy.

They are doing a piece-rate job that is highly future-oriented yet extremely primitive—producing the data feed required by AI giants in Beijing, Shenzhen, and Silicon Valley to build large models.

No one thinks there’s anything wrong with that.

A new production line on the Loess Plateau

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

Autonomous driving needs to recognize traffic lights and pedestrians; large models need to distinguish 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 after it has ingested tens of millions of images will it learn to recognize on its own.

This job doesn’t require a high education level—only patience, and an index finger that can click endlessly.

In the golden age of 2017, a simple 2D box could cost more than a single dime; even companies offered a high price of 5 dimes. Labelers with fast hands could work more than ten hours a day and earn five or six hundred yuan. In county towns, this is definitely considered a high-paying, respectable job.

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

By 2023, the unit price for simple image labeling had already been smashed down to 3 to 4 cents. The drop exceeded 90%. Even for 3D point-cloud images with higher difficulty—images made of dense points that require magnifying countless times to see the edges clearly—labelers must still draw a 3D box in three-dimensional space that includes length, width, height, and rotation angles, tightly and precisely wrapping the vehicle or pedestrian. Yet even such a complex 3D box is only worth 5 cents.

The direct consequence of the unit price collapse is a surge in labor intensity. To tightly hold onto the meager base salary of two to three thousand yuan per month, labelers must constantly, without stopping, improve their hand speed.

This is absolutely not an easy white-collar job. In many labeling bases, management is strict to the point of suffocation. During work, phone calls are not allowed, and mobile phones must be locked in storage compartments. The system will accurately record each employee’s mouse trajectory and dwell time. If they stop for more than three minutes, back-end warnings will come like whips.

What’s even more crushing is the tolerance rate. The industry passing threshold is usually above 95%; some companies even require 98%–99%. This means: if you draw 100 boxes and make only 2 mistakes, the entire image will be rejected and sent back for rework.

Motion images are continuous frames. If a vehicle changes lanes, it will be occluded, and labelers must rely on inference to find each one out; in 3D point-cloud images, as long as an object has more than 10 points, you must draw a box. In a complicated parking-spot project, if the line is drawn too long or you miss a label, quality checks will always pick out the problems. Getting a single image sent back for rework four or five times is commonplace. In the end, after spending an hour, you might only get a few cents.

A labeler in Hunan posted her settlement sheet on a social platform. After a full day’s work, she drew more than 700 boxes, with a unit price of 4 cents, for a total income of 30.2 yuan.

It’s an extremely bifurcated picture.

On one side, tech bigwigs look dazzling at press conferences, talking about how AGI will liberate humanity; on the other, in county towns on the Loess Plateau and in the mountains of southwest China, young people stare at their screens for eight to ten hours every day, mechanically drawing boxes—thousands, tens of thousands—and even at night they dream, with their fingers drawing lane lines in midair.

Someone once said that AI’s exterior is a luxury car roaring past, but when you open the door, you find that inside, there are a hundred people riding bicycles, clenching their teeth and desperately pedaling.

No one thinks there’s anything wrong with that.

A piece-rate job to “teach machines how to love”

Once the bottleneck of image recognition is broken through, large models enter a deeper stage of evolution: they need to learn to think, converse, and even show “empathy” like humans.

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

Put simply, it means having real people score the answers generated by AI, telling it which answer is better and more aligned with human values and emotional preferences.

The reason ChatGPT looks “like a human” is because behind it there are countless RLHF labelers giving it lessons.

On crowdsourcing platforms, these labeling tasks are often priced openly: 3 to 7 yuan per item. Labelers need to provide extremely subjective emotional scores for the AI’s answers—to judge whether this answer is “warm,” whether it “shows empathy,” and whether it “takes care of the user’s emotions.”

A bottom-layer worker who holds a monthly salary of two or three thousand yuan, hustles desperately in the real-world mire, and even has no time to care about their own emotions—yet in the system they must serve as an AI’s emotional mentor and a judge of values.

They need to forcefully crush those extremely complex and subtle human emotions—warmth, empathy—into cold scores from 1 to 5. If their scoring disagrees with the system’s pre-set correct answers, they will be judged as failing the accuracy requirement, and their already meager piece-rate wages will be deducted.

This is a kind of cognitive hollowing-out. Human emotions—complex and hard to parse—the morality and compassion, are being forced into the funnel of algorithms. In the cold quantized and standardized scales, they are squeezed dry of the last bit of warmth. When you marvel that the cyberbeast on the screen has already learned to write poems, compose music, and inquire about warmth and coldness, even putting on a guise of melancholic sensitivity; outside the screen, that group of originally vivid humans, in day-after-day mechanical judgment, degenerates into emotionless scoring machines.

This is the most hidden side of the entire industry chain. It has never appeared in any financing news or technical white papers.

No one thinks there’s anything wrong with that.

985 master’s degrees and youth from small towns

The bottom-layer box-drawing work is being crushed by AI’s caterpillar tracks. This cyber production line is spreading upward, starting to consume higher-level cognitive labor.

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

On major recruitment platforms, a special type of part-time job has begun flashing frequently, such as “large-model logic reasoning labeling” and “AI humanities training instructor.” The entry requirements are extremely high, often requiring “a 985/211 master’s degree or above,” and involving specialized fields such as law, medicine, philosophy, and literature.

Many graduate students from prestigious universities are attracted and pour into these outsourced groups at big tech companies. But they quickly find that this isn’t really a relaxing cognitive gym—it’s a mental torture.

Before they can take orders officially, they must read dozens of pages of scoring dimensions and evaluation criteria documents, and complete two to three rounds of trial labeling. After passing, in official labeling, if the accuracy is below the average level, they lose eligibility and are kicked out of the group chat.

Most suffocating of all is that these standards aren’t fixed at all. When facing similar questions and answers, if they score using the same way of thinking, the results may be completely opposite. It’s like doing an exam that is never finished and has no standard answers. You can’t improve accuracy through effort or study—you can only endlessly spin in place, consuming brainpower and physical energy.

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

Knowledge, once seen as the golden ladder that breaks barriers and helps people climb upward, now has fallen into becoming digital fodder for algorithms, increasingly more complex to chew. Under the absolute power of algorithms and systems, the 985 master’s degree holders in ivory towers and the youth from small towns on the Loess Plateau are facing the most bizarre convergence of different paths toward the same end.

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

It’s the same overseas. In 2024, Apple directly cut an AI voice labeling team of 121 people in Santiago. These employees were responsible for improving Siri’s multilingual processing ability. They had 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.

In the eyes of tech giants, whether it’s the box-drawing aunt in a county town or the logic training instructor who graduated from a top university—at essence, they are all “consumables” that can be replaced at any time.

No one thinks there’s anything wrong with that.

A trillion-branch Babel, piled with blood sweat worth a few cents

According to data released by China’s Academy of Information and Communications Technology, in 2023 the size of China’s data labeling market reached 6.08 billion yuan; by 2025, it is expected to reach 20–30 billion yuan. It is predicted that by 2030, global sales in data labeling and services will skyrocket to 117.1 billion yuan.

Behind these figures are value-adding frenzies among tech giants like OpenAI, Microsoft, and ByteDance, with valuations worth hundreds of billions to trillions of dollars at the drop of a hat.

But none of this staggering wealth flows to the people who truly “feed” AI.

China’s data labeling industry displays a typical inverted pyramid outsourcing structure. At the very top are tech giants that tightly hold the core algorithms; the second tier is large data service suppliers; the third tier is data labeling bases and small-to-medium outsourcing companies distributed across the country; at the very bottom are those labelers who get paid by the piece.

Every layer of outsourcing has to harshly skim off a layer of oil and water. When the unit price set by the big company is 5 dimes, after layers of stripping and extraction, what reaches the labelers in county towns may be less than even 5 cents.

In his book “Technological Feudalism,” Yanis Varoufakis, former Greek finance minister, put forward an insight with great penetrating power: today’s tech giants are no longer capitalists in the traditional sense, but “cloud lords” (Cloudalists).

What they own is not factories and machines, but algorithms, platforms, and computing power—these are digital territories in the cyber age. In this new feudal system, users are not consumers but digital tenants. Every like, comment, and view on our social media is freely offering data to the cloud lords.

And those data labelers scattered in the lower-tier markets are the lowest-level digital serfs in this system. They not only have to produce data, but also have to clean, classify, and score massive volumes of raw data, converting it into high-quality feed that large models can digest.

This is a covert movement of cognitive territory-taking. Just like the enclosure movement in 19th-century Britain drove farmers into textile factories, today’s AI wave drives those young people who can’t find a place in the real economy to the front of screens.

AI has not smoothed away the class divide; instead, it has built a “data and sweat conveyor belt” that runs from China’s central and western county towns straight to the headquarters of tech giants in Beijing, Shanghai, Guangzhou, Shenzhen. The narrative of technological revolution is always grand and dazzling, but its underlying color is always large-scale consumption of cheap labor.

No one thinks there’s anything wrong with that.

A future no longer requiring human beings

The most brutal ending is coming soon—faster and faster.

As large models’ capabilities leap, those labeling tasks that once required humans to labor day and night are being taken over by AI itself.

In April 2023, Li Xiang, founder of Ideal Auto, disclosed data on a forum. Previously, Ideal would need to manually label autonomous driving images for roughly 10 million frames per year. Outsourcing costs were close to 100 million yuan. But when they used large models for automated labeling, what used to take a year could basically be done in about three hours.

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

A saying that circulates in the industry is extremely true, mocking itself: “As much intelligence as there is, there is as much human labor.” But now, big companies’ investment in outsourcing data labeling has already shown a cliff-like decline of 40%–50%.

Those young people from small towns who spent countless days and nights hunched in front of computers, burning their eyes until they turned bloodshot—fed a huge beast with their own hands. But now, this beast is turning around and smashing up their jobs.

As night falls, the office buildings in Pingcheng District of Datong remain as pale as day. Young people at shift handover silently exchange their exhausted shells in the elevator lobby. In this folded space sealed in by countless polygon boxes, no one cares what kind of world-historical leap the Transformer architecture across the ocean has just taken next, and no one can make sense of the roaring computing power behind hundreds of billions of parameters.

Their gaze is only welded to the red-green progress bar in the back end, representing the “passing line,” calculating whether those few points and a few cents of piece-rate numbers can be pieced together into a decent life by the end of the month.

On one side are the ringing of the bell at Nasdaq and endless coverage by tech media, with the giants toasting to the arrival of AGI; on the other side are these digital serfs who fed AI one mouthful at a time with their own flesh and blood, who can only wait in anxious, aching dreams for that huge beast they personally raised to, in some seemingly ordinary morning, casually kick their jobs away.

No one thinks there’s anything wrong with that.