This is part one of a two-part series on China's annotation workers.

When the public talks about AI today, the conversation is usually grand and abstract. Will it replace you, if not humanity? Are our ethical guidelines strong enough, or will it spiral out of control? And when people picture those who work in AI, they tend to imagine a class of brilliant elites: men and women fought over by global capital, whose ideas may, indeed, alter the course of civilization.

But when I think of AI, a different image comes to mind.

It is a little after nine in the morning. A group of young mothers has just dropped their children off at kindergarten and primary school. They turn and hurry to work. Once inside the computer lab, they barely have time to catch their breath before logging in and scrambling to claim the latest batch of tasks — image-labeling assignments for autonomous vehicles, freshly posted by one of China’s tech giants.

This is the world that Xia Bingqing, of East China Normal University, and I have spent five years studying. AI, after all, must be fed. It must practice. It must be taught, piece by piece, the knowledge, values, and norms of human society. In that sense, AI is not nearly as futuristic, or as elite, as its mythology suggests. It does not hover in the clouds. It rests on the ground — specifically, in the valleys of inland China, inside “data-labeling centers” built in relocation communities created through poverty-alleviation campaigns. It lives at rows of workstations, in mice and headsets, in timers, correction slips, and rework orders.

The workers Xia and I study have been described in Chinese media as “the teachers who train AI.” That is not exactly wrong. But it is too romantic. A more accurate description would be this: They perform the labor of translating the world into a language machines can digest. They take raw materials — images, audio, text — and sort, classify, and tag them into training data that algorithms can learn from, correct against, and iterate upon. They draw boxes around people, cars, and potholes in photographs. They slice recordings into alignable words and phrases. They score conversations, telling the machine which responses conform to human values and social judgment, and which do not.

Internationally, this kind of data labor is often discussed through the language of global division: Engineers and models are concentrated in the Global North, while annotation is outsourced to the Global South. Pay is measured in cents per task or a little more than a dollar an hour. The business is brutally simple. Work goes where labor is cheapest.

China’s internet giants, however, did not simply move this work offshore. To be sure, these firms also care about cost. But in many sensitive projects, they fear something else even more: leakage. Data, for them, is not just fuel for training. It is closer to a research-and-development roadmap. So they adopted what might be called an inland-sourcing model, moving labeling work away from municipal headquarters in Beijing, Hangzhou, and Shenzhen into inland provinces and regions such as Shanxi, Shaanxi, Xinjiang, Guizhou, and Henan, where they own or tightly control their data-labeling centers.

These less-developed inland regions welcomed the companies with open arms. For many smaller cities, participation in the AI economy seemed an impossible aspiration. Compute power, technical talent, venture capital — none of these are readily available. But the “data industry” was attainable. It had a lower barrier to entry. It is labor-intensive. In practice, it resembles factory-line work far more than anything people usually associate with cutting-edge technology: the same density, the same low unit price, the same attrition, the same damage to the eyes and the nerves. It does not require advanced credentials or rarefied insights. It requires diligence, endurance, and patience. For local governments, landing such work meant something even more attractive: It could be written into polished reports as digital-economy employment, AI-related job creation, industrial growth, and community stabilization. It could become a piece of political achievement.

That mattered all the more because many of these regions were losing population. In the relocation communities we studied, several villages had been moved down from the surrounding mountains and resettled together. The phrase one community secretary repeated most often was: “We have to keep people here.” This was not a slogan. It was a cluster of concrete problems. People had moved down the mountain, but their land was gone. Their old skills had little value. Young adults were leaving. The elderly and children remained behind in apartment blocks. If people were truly to settle here, they needed jobs — not necessarily lucrative, but reliable; not necessarily prestigious, but nearby.

Then, in 2018, an opportunity arrived. The secretary’s community began negotiating with a major technology company, which I will call B-Tech. After several rounds of talks, B-Tech placed its first data-labeling center in the valley. The relocated community offered three years of rent-free space, followed by subsidies. The community took responsibility for utilities and network maintenance. The company, in turn, promised to provide jobs gradually and, at the community’s insistence, to prioritize hiring “women in difficult circumstances” — those with less education, those who were middle-aged, and those with child-care responsibilities.

From a distance, this arrangement looked like a neat story of mutual benefit: a win-win partnership between coastal tech firms and underdeveloped inland regions. Up close, the road was far rougher.

The legal representative of these centers was often the community secretary himself. From within the community, he would pick “the young person who had studied best” to serve as center manager, while workers were drawn from the local labor pool. The stickiness of community and kinship networks seemed, at first glance, to guarantee the stable labor force the tech firms wanted. But this stability was never gentle.

The problem was obvious. Labeling orders from big tech came in waves. At times they surged; at times they dried up. When the trough came, morale was the first thing to loosen. If there was no work and wages became unstable, people left. Once they left, the next peak required recruitment, retraining, and a fresh round of adjustment. Quality fell. Rework increased.

Local governments stepped in with an improvised but revealing solution: They used training programs and subsidies to keep workers in a state of readiness, turning the local economy itself into a shock absorber for the platform. When orders were scarce, centers sent workers to employment-bureau training sessions that, in substance, were really just basic labeling drills. Anyone who showed up and signed in received a daily subsidy of 50 yuan ($7). When orders dropped sharply, centers sought additional subsidies associated with poverty-alleviation workshops, allowing eligible workers to receive support up to 500 yuan. On paper, this was employment assistance. In practice, it served a quieter function: suppressing the impulse to quit and preserving a labor pool in a permanently usable state.

Daily management inside the centers was shaped by local knowledge. Managers knew whose child got out of school at what hour, whose eyes had recently become inflamed, whose mother-in-law was bedridden, who could work late but had to return to the village on weekends. During peak periods, workers were divided into groups. Those considered younger, stronger, less burdened by family obligations, and more efficient were assigned to the high-output teams. Mothers entangled in domestic responsibilities were typically pushed toward simpler tasks.

One project manager responsible for AI-data operations at a major tech company told me, with unusual candor, that the company had long assumed algorithms could generate a more precise, granular, and inescapable system of control. They had spent years developing optimization algorithms to produce division-of-labor and management strategies. But doing that required, in his words, “a tremendous amount of analytical work, data, and compute power.” Put plainly, it required “a tremendous amount of money” — starting in the tens of millions of yuan. In the end, they discovered that the task allocation and management done by local center managers, drawing on experience and intimate knowledge, was “far more accurate and effective than algorithms — and much cheaper.” Indeed, the accuracy rate at company-owned centers could reach 97% to 98%, clearly outperforming most external platforms and third-party factories.

Many people imagine the story of AI unfolding in the cloud — in compute, models, and parameters. But in China’s remote valleys, it looks more like a production line crawling along the ground. Every click, every correction, every hurried sprint out of the center lab at 4:30 p.m. to pick up a child is part of the force that moves it forward. And once we lower our gaze, we begin to see that intelligence does not belong to machines alone. It is also built from the labor of those people hardest to see.

Portrait artist: Wang Zhenhao.

(Header image: Visuals from Amy DeVoogd and Shijue Focus/VCG, reedited by Sixth Tone)