Just after midnight in a Hangzhou, China livestreaming studio, a woman sells skincare to an audience of strangers. Her name is Amber-1. She blinks, tilts her head, and holds eye contact a beat too long as she works through her pitch in polished, tireless Mandarin: “This cream is what every girl dreams of. This is our star product. Its ingredients are…” She will keep going until sunrise, because Amber-1 is an AI avatar, and she never needs to sleep.
Viewers notice. “Robotic,” one writes. “Fake,” “AI,” others chime in. They are right, and in this studio it does not matter. By morning, Amber-1 will have sold roughly 60 percent of what her human predecessor sold in the same overnight slot, at a fraction of the cost. Within months, the agency that runs this channel will cut its roster of human streamers from six to two.
A few miles away, another company tested a similar avatar, this one modeled on its star human host. Viewers there called it fake too, but this time the verdict impacted sales. Watch time plummeted, return rates were cut in half, and the experiment ended almost as soon as it began. The machine was moved backstage, and the human hosts kept their jobs.
Why was the avatar viable in one studio and unacceptable in another?
In a recent study of China’s live-commerce industry, based on fifteen months of ethnographic fieldwork and 141 interviews, I followed AI avatars into these two companies as they adopted identical technology yet experienced opposite outcomes. The answer, I found, had little to do with the machines. What separated the two studios was how each measured human work. One had built an evaluation system that overlooked the emotional labor its streamers performed; the other built a system capable of valuing it. That difference in measurement rendered one group of workers disposable and made the other appear irreplaceable.
I call these systems of measurement authenticity regimes: the metrics, stories, and workplace rules through which an organization defines what counts as an authentic emotional performance. Every workplace has one, whether anyone names it or not. It decides which parts of a job become visible and valued, and which parts fall outside the frame.
Authenticity regimes let us see something new about an old problem. Since Arlie Hochschild’s The Managed Heart, we have known that emotional labor is governed by “feeling rules.” But for decades, those rules were enforced by human judgment: a supervisor who watched a service worker smile, a customer complaint or compliment that shaped worker awareness of the impact of their expressions. Whether the warmth was “real” was an interpretive call. Platform work changes the enforcer. On livestreams, the authenticity regime runs through dashboards and are determined by watch time, return rates, conversion so that “realness” is no longer something a manager perceives but something the numbers register. Whatever moves the metrics is authentic. What organizations certify as “authentic” turns out to be a rules-based order rather than any deeper truth about feeling.
These rules shape workers’ own sense of what is real. During a stream I observed, a streamer teared up on camera at precisely the moment her team had scheduled an emotional peak. “I cried on cue,” she told me afterward, “but also, I did feel it. That’s why it works.” Another streamer who replayed her best-performing moments each night like game tape put her affective adjustments in starker terms: “The chart tells you when you’re being authentic.” For these workers, genuine feeling and metric calibration had become inseparable. The regime does more than evaluate performers; it retrains them, teaching them what their own sincerity should feel like.
When authenticity was a matter of human judgment, the question of whether a machine’s affect equaled a person’s had no real purchase. Once a given regime converts emotional performance into metrics, humans and machines occupy the same dashboard questions of authenticity become a matter of calibration. For humans and AI avatars alike, the fate of emotional labor in the AI age depends less on what they can do than on the regime evaluating them. Identical avatars will meet opposite fates in workplaces that measure differently.
The two business models I studied built distinctive authenticity regimes long before AI arrived, with different ways of measuring appropriate emotions. In one, hosts built personal followings among viewers who returned, night after night, for a felt connection to a person. In this model, managers tracked whether viewers came back, lingered, commented, and joined fan clubs, because those numbers were selected as the standard for real connection. When one company tested an avatar modeled on its star host, the dashboard registered the failure immediately with watch times and visits falling. In the end, the avatar was pulled.
In contrast, what I call the “brand streaming regime” used transaction metrics, clicks and sales, to evaluate the effectiveness of hosts. When avatars were deployed, complaints of inauthenticity were ignored if sales volume was maintained.
The stakes in these parameters of assessment are of great consequence for the future of employment. Ultimately, the system of measurement defines acceptable performance, and AI avatars who meet these metrics can readily replace their human counterpart. Workers can lose their jobs when machines learn to fully replicate their performance, but AI may also replace humans as firms shift their expectations for those performances. The oft-repeated claim that machines will never truly feel, connect, or care may hold, and yet the response, increasingly, will be those systems of assessment shift to make AI performances acceptable.
In the struggle over replacement, workers and unions should bargain not only over wages and hours, but over the metrics that measure their performance. If scholars have documented how emotional work is chronically undervalued and often invisible to the systems that evaluate work, in the age of AI, formalizing evaluation in the form of metrics becomes the mechanism of replacement Protecting human work will require building evaluation systems that can see and value what people actually do. Otherwise, we will keep telling ourselves a flattering story about machines that simply won, when the real decision happened earlier and more quietly, inside a dashboard, when someone concluded that the part of the job they never counted was never worth counting.
Read More
Jun Zhou. “Constructing Replaceability: Authenticity Regimes and the Automation of Emotional Labor” in Work & Occupations, 2026
About the Author
Jun Zhou is a doctoral candidate in Sociology at the University of Michigan, with a graduate certificate in Science, Technology, and Society, and an ACLS/Mellon Dissertation Innovation Fellow. She is a scholar of labor, gender, and political economy, using historical ethnography and social theory to examine how technological transformations reorganize gendered labor and reshape the futures women can plan.