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

At 4:30 in the afternoon, after school pick-up, Yingzi brings her daughter back to her workstation. She tells the girl to start her homework, then turns back to the screen.

Her task that day is to annotate dense pedestrian traffic captured by shopping-mall surveillance cameras. One hundred frames of video. Hundreds of human figures. Each body must be broken down into head, torso, legs, and arms, then boxed and labeled inch by inch so the machine can learn to see a person as a person.

Yingzi pulls the monitor so close it is almost pressed to her face. She has done this work for only two or three years, and already her worsening myopia forces her to squint, carving a deep vertical line between her brows. If one of her boxes is off, the team leader will tag her directly alongside the faulty image in the company’s chat group, correcting her in public without the slightest pretense of tact. Meanwhile, her daughter, bent over Chinese homework beside her, interrupts every so often to ask how to pronounce an unfamiliar character.

Yingzi works in one of the data-labeling centers Xia Bingqing, of East China Normal University, and I studied over five years. As China’s AI industry has grown, more and more centers like this one have cropped up in underdeveloped inland regions, taking on labeling work outsourced from coastal tech giants.

In European and American scholarship, the workers at the bottom of the digital economy often appear as “ghost workers”: scattered across the globe, taking orders in the cloud, submitting tasks, and vanishing again. Platforms see only IDs, ratings, and completion rates. The person herself disappears behind the metrics. And it is precisely this invisibility that often draws women in. Platform “flexibility” fills fragments of spare time between care work and housework, turning leftover labor into supplemental income.

In China’s data-labeling centers, most workers are also women — and specifically mothers like Yingzi. The difference is that they do not take orders at home. They clock in at the center and sit in a standardized computer room under direct supervision.

For these women, the line between work and life is barely a line at all. Many spend their lunch break running to buy groceries, then hurry home to cook lunch and prepare dinner before racing back to their desks. At 4:30, they leave again to pick up their children, only to return later and put in another hour or two. Some simply bring the child back to work and keep drawing boxes while helping with homework. Their time is sliced wafer-thin and pasted into the cracks of daily life.

It is precisely within this arrangement that management becomes difficult. One center manager, nicknamed Heizi, is well educated but still young. In kinship terms, he belongs to the “nephew” generation relative to the mothers he supervises. He once tried posting an attendance chart and announcing that everyone would now strictly observe work hours. Before the rule could settle, more than 20 “aunties” surrounded him and lectured him: You’re not even married. You don’t understand how hard real life is. The attendance chart remained on the wall, ignored as if it were decorative paper.

The key point here is not that women are somehow undisciplined. It is that discipline itself that is rewritten by family structures. The force that truly governs these mothers is not the manager, but the household. The daughter-in-law is obliged to keep earning for the family while still treating paid work as secondary to domestic duties, especially her role of being filial to her mother-in-law on her husband’s behalf. Whatever she says goes.

This power dynamic is on display outside of the labeling centers. Mothers-in-law often gather in the small plaza — much as they once did in the village square back in their mountain villages before relocating — chatting in the sun with an air of unhurried authority, like an invisible center of domestic power.

Weighing on these women as well is the moral pressure of motherhood, which compels them to divide their attention even during the busiest work periods. When these women were children themselves, their parents left the villages for better-paying coastal jobs, and could not take their children with them due to restrictive household registration policies. They grew up as “left-behind children.” Now, these mothers do whatever they can to make sure their children do not suffer the same emotional absence.

Local governments understand perfectly well how tied this generation of rural women is to their parents-in-law and their children. The point of creating data-labeling centers was not just employment; it was to keep residents from leaving the resettlement communities. In local policy practice, demonstrating stability and making relocation stick often counts more than job creation alone. And because patriarchal customs dictate that men should migrate outward for better-paying work, women became the crucial population for anchoring households in place. Once women stayed, the apartment remained occupied, the elderly were looked after, the children stayed in school, and the family remained rooted in the community.

That is why local governments insisted, during negotiations with the tech firms, that women should be hired first. Later, some of these centers were granted the title of “Heroine’s Workshop.” Outstanding female workers are regularly pushed before cameras to testify to the dignity of “working close to home.” And these mothers did, in fact, work hard. They volunteered for overtime, practiced their skills, chased higher accuracy and speed, and tried not to betray the honor of being held up as “heroines.”

But here the story requires a sharper reading. These women have not been liberated from the household. On the contrary, the “women’s empowerment” narrative surrounding the centers often binds their domestic burdens more tightly to the labor regime, though in gentler language. One example is the so-called “4:30 classroom,” where children can be watched after school until their mothers get off work. It looks like benevolent infrastructure. It is also a clear message: You may work, but you must simultaneously remain responsible for mothering.

Beneath this moral binding of care lies a political economy of guilt. Many of these mothers had once left home to work in distant factories and, in doing so, missed years of their children’s lives. After returning, that absence hardened into a form of compensation: They would sacrifice sleep, leisure, even wages if it meant keeping the child physically near. In our interviews, one mother described it plainly. She had spent seven years away. After coming back, she resolved that “I have to raise my child myself.” When her daughter’s grades began to slip, she fought for an earlier shift so she could free up more time for supervising her daughter’s homework.

That guilt produces a distinctive form of labor compliance. Many are more willing to accept the logic of “It’s fine if I earn a little less,” because the job is defined, in their own minds, as a choice that keeps them close to home and able to care for others. So when the center reorganizes labor to “hold onto orders,” diverting the more profitable tasks toward a small high-efficiency group, many mothers do not openly protest when they are assigned the less desirable work. They are more likely to interpret the difference as the consequence of their own choice: “I have to care for my child, so of course I can’t compete for the harder, better-paying tasks.” Over time, the halo of the “heroine worker” becomes a lubricant. It makes structural inequality easier to swallow. It turns “you have been relegated to a secondary position” into “you nobly chose your family.”

There is another reason centers prize mother-workers: Data labeling requires precisely the steadiness, endurance, and tolerance for tedium that women are so often forced to cultivate through life itself. Technical breakdowns are constant. Internet outages are routine. A worker may be on the verge of finishing a task package when the connection drops. The data stops loading. Previously completed work may be wiped. Once the system comes back online, much of it may have to be redone from the start. Younger male workers in the room often become impatient first — chatting, scrolling, drifting. Older women are more likely to remain fixed in place. Some say nothing and stare at the unresponsive page, refreshing it again and again until their hands ache. Others turn on their own mobile hotspots, burning through personal data just to keep working.

Yet even this hard-won reputation for diligence and resilience is losing its value. As large language models rise, the threshold for data-labeling work is climbing. The smarter the models become, the more skill is demanded of the people teaching them how to understand the world. Major tech firms now pressure local centers to recruit labelers who better “match” the cognitive level and learning speed of large models. Who counts as “better-matched”? Local managers often cannot define it clearly. And on the ground, managers increasingly describe women — and mothers in particular — as less suited to the work’s rising demands. Women are described as less comfortable with computers, while mothers are seen as too stretched by family responsibilities to devote the time, energy, and concentration needed to upgrade their skills.

And so the same women who, yesterday, stood beneath the sign for the “Women’s Poverty-Alleviation Workshop,” praised as careful, steady, hardworking, and exemplary symbols of employment close to home, are today judged “mismatched,” “unskilled,” “unable to keep up,” and “unsuitable.”

The pattern is an old one. When an industry is struggling uphill and needs labor that is patient, meticulous, and cheap, women’s endurance becomes indispensable. When that same industry rises, consolidates, and seeks prestige, women suddenly reappear as burden, flaw, residue — as disposable material, used and then cast aside.

Perhaps Yingzi will leave one day, as many women like her eventually do, without ever being told what, exactly, changed. One day, she is praised as a “super mom,” held up as proof that a mother can train AI and raise a child at the same desk; the next, she is judged too distracted, too slow, too unskilled for the very future of AI her labor helped build.

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