The manufacturing of (Un)stability: Possible but improbable stability in long-term content moderation on Chinese social media platforms

Introduction

At 1 p.m., Don returned from his lunch break, not to resume frontline moderation tasks, but to train six new colleagues in moderation standards. Seven years earlier, Don had arrived in Tianjin, often dubbed the “content moderation capital of northern China” (The Paper, 2022), as a recent graduate who, like them, stumbled into content moderation at Company A with limited options. Since then, through internal transfers, he has rotated through four roles, from first-line content moderator to quality inspector, within the moderation system. Don’s story is not an exception. Rather, it exemplifies a broader shift in China’s content moderation industry, where what was once treated as precarious labor (Gillespie, 2018; Roberts, 2019; Zhai, 2023) is becoming institutionalized, specialized, and embedded within state and platform governance structures.

The last two decades have witnessed the deep engagement of digital platforms in everyday life, leading to an unprecedented expansion of platform-based information, commonly referred to as “content” (Steinberg, 2019). This development has given rise to a range of occupations centered on content-based information production, including content creators (Craig et al., 2021; Lin, 2023; Zeng and Kaye, 2022), data annotators (Fu et al., 2025; Wu et al., 2025), and content moderators (Li and Zhou, 2024; Xiao, 2024; Zhao and Zhang, 2024). They perform different roles in production, dissemination, and moderation, forming a division of informational labor production chain around digital platforms and having become the protagonists of the Internet’s dream-making myth (Xia, 2021). However, content moderation “as the multi-dimensional process through which content produced by users is monitored, filtered, ordered, enhanced, monetised or deleted on social media platforms” (Badouard and Bellon, 2025: p.3) are positioned at the end of this chain, yet they remain an essential component of platform governance, as they contribute to balancing hate speech and misinformation with the promotion of online democracy and freedom of expression on platforms and community forums (Jørgensen and Zuleta, 2020). China has the world’s largest Internet user base, with over 1.1 billion users as of 2024 (CNNIC, 2024). New formats such as social media, short video, and livestream e-commerce have proliferated, creating a highly complex content ecosystem. This vast scale, combined with heightened political sensitivity and evolving regulatory demands, exerts continuous pressure on platforms to refine their governance mechanisms.

Platforms not only avoid disclosing the operational systems behind content moderation but often construct the illusion that moderators do not exist (Roberts, 2018, 2019) Their work, obscured by opaque rules and ambiguous mechanisms, has been described as a “black box” (Ozanne et al., 2022), “invisible labor” (Roberts, 2019), or “ghost work” (Gray and Suri, 2019). Bound by non-disclosure agreements and social stigma, moderators remain low-profile, serving as invisible custodians of platform governance (Myers, 2018; Roberts, 2016, 2019; Zhai, 2023). Lacking employment stability and social protection, commercial moderators are often classified as precarious workers (Kalleberg, 2009), a reality widely acknowledged in current scholarship (Gillespie, 2018; Roberts, 2018, 2019; Zhai, 2023).

However, as moderation becomes deeply embedded within platform operations and state governance, analytical frameworks that frame content moderation as a form of precarious labor warrant critical re-examination. Content moderation has consequently shifted from a peripheral and reactive task into a core infrastructural function, integral not only to platform operations but also to broader systems of information control and state governance. In 2023, the content moderation service market in China reached a value of 9.094 billion RMB (1.273 billion USD), with a conservative estimate of more than 100,000 people employed in this sector (Gelonghui, 2024). Renrui Human Resources (2025), the largest content moderation outsourcing company in China, reports that among its 41,868 gig workers, the monthly turnover rate in 2024 was approximately 6.8%, a decrease from 8.3% in 2023. In contrast, the 2024 Employee Turnover and Salary Adjustment Survey Report published by 51job (China's largest recruitment platform) indicated that the overall employee turnover rate across Chinese enterprises reached 15.3%, with the Internet/tech sector recording a significantly higher average turnover rate of 22.7% in 2023 (Smartsalary, 2025). In this context, content moderators seem more stable than often assumed.

This paradox prompts critical questions about how stability is produced in a sector widely regarded as precarious. In China’s food-delivery sector, for example, platform work has become “sticky labour,” as de-flexibilisation is produced through algorithmic management, staffing arrangements, and the cultural normalization of platform-dependent precarity (Sun et al., 2023), also referred as the “good bad” job (Cameron, 2024). Similar patterns have been observed elsewhere; in Finland, for example, long-term moderators benefited from stable incomes and social protections, advocating for the re-humanization of platforms (Ruckenstein and Turunen, 2020). Other studies have explored the motivations and career trajectories of long-term gig workers (Jabagi et al., 2019; Ray et al., 2024). A sole focus on precarity fails to explain why content moderation continues to grow in scale while maintaining relatively low turnover. Understanding this institutionally produced form of “stability” is thus crucial to addressing the broader dynamics of labor, control and governance in digital economies.

In what sense can content moderators be considered “stable”? In this study, “long-term moderators” are defined not only by experience but also by legal and organizational criteria. Under China’s Labour Contract Law (LCL) (Labour Contract Law of the People’s Republic of China, 2012), fixed-term contracts typically span 1–3 years, and most platform companies offer 3-year contracts with potential renewal. While the average tenure remains around 7 months (Huxiu, 2022), our focus is on those who have renewed at least once and worked for over 3 years, often rotating across multiple roles within the moderation system. Examining this group reveals a more complex ecosystem in which the tension between stability and mobility is not incidental but actively structured through organizational mechanisms and governance demands. This structured tension is a key feature of how “stability” is manufactured in China’s content moderation industry.

This article contributes to digital labor studies and critical labor process theory by conceptualizing “manufacturing stability” as a key mechanism through which long-term content moderation is produced and sustained in China. By integrating micro-level adaptation strategies, meso-level organizational structures, and macro-level governance imperatives, it shows how long-term moderation emerges not from improved labor conditions but from the intersection of individual survival strategies and structural disciplining forces. This perspective challenges idealized assumptions of long-term stability as decent work and demonstrates how content moderation in China is institutionally shaped and politically embedded, offering new insights into the reconfiguration of labor and governance in the platform economy.

Literature review

Data and method

This study employs qualitative research methods, primarily semi-structured in-depth interviews, to explore the experiences and practices of long-term content moderators in China. The research received ethics approval from the authors' institutional ethics committee prior to data collection. The main data collection phase was conducted from July 2024 to May 2025. Ethical considerations were rigorously adhered to throughout the research process. Participants were fully informed about the objectives, scope, and implications of the study, retaining the right to withdraw at any stage without any adverse repercussions.

Fieldwork and data collection

Fieldwork for this study leveraged professional networks established through the first author’s prior employment as an account operations specialist at a leading social media company (May–October 2022) in Tianjin ¹ , a city whose high concentration and representativeness make it an ideal fieldwork site for examining platform labor, technological practice, and state governance.

This study utilized a purposive and snowball sampling approach (Parker et al., 2019), progressively identifying participants across diverse roles and experience levels in the moderation industry—from frontline moderators to supervisory and managerial staff. We conducted semi-structured interviews with 21 participants, including moderators, supervisors, quality inspectors, risk control specialists, trainers, and managers. Interviews lasted 50–130 minutes and explored themes such as job stability, adaptation strategies, workplace dynamics, career opportunities, and professional trajectories. All participants gave informed consent, understood their confidentiality and anonymity rights, and agreed to audio recording. Identifying details were anonymized in all publications and presentations. See Table 1 for a summary of interviewee details.

OPEN IN VIEWER

Table 1.

Sample description (N = 21 interviews).

Nr.	Gender	Work experience	Professional role	Platform	Forms of employment
#1	Male	7 years	A-B-C-D	Platform A	Insourcing
#2	Female	4 years	A-C	Platform A	Insourcing
#3	Female	3.5 years	A	Platform A	Insourcing
#4	Female	6 years	A-E	Platform B	Formal Staff
#5	Male	3.5 years	B	Platform A	Insourcing
#6	Male	3 years	B-C	Platform A	Formal Staff
#7	Male	4 years	B-A	Platform A	Insourcing
#8	Male	4 years	A	Platform C	Insourcing
#9	Male	5 years	A-C-F	Platform D	Outsourcing-Insourcing
#10	Male	5.5 years	A-F-E	Platform E	Outsourcing
#11	Male	3.5 years	G-A	Platform F & Platform G	Insourcing-Outsourcing
#12	Female	4.5 years	A-E-H-F	Platform F	Insourcing
#13	Male	4 years	A-H-E	Platform A	Insourcing
#14	Male	10 years	B-I	Platform H & Platform A	Insourcing-Formal Staff
#15	Male	4 years	A-E-F-C	Platform F	Insourcing-Formal Staff
#16	Female	5 years	A-E	Platform F & Platform I	Outsourcing-Insourcing
#17	Male	4 years	A-F	Platform S	Insourcing
#18	Female	5 years	A-E	Platform S	Insourcing
#19	Female	3 years	A-F	Platform J	Formal Staff
#20	Male	3.5 years	A	Platform K	Insourcing
#21	Female	4 years	A-C	Platform L	Formal Staff

The “Professional Role” column uses the following codes:

A: Front-Line Content Moderator

B: Senior Content Moderator

C: Content Security Operation

D: AI Data Annotator

E: Quality Inspector

F: Supervisor

G: Labeling

H: Risk Control

I: Data Analyst for Content Moderation.

Data analysis

All interviews were digitally recorded and transcribed verbatim. These transcripts formed the primary dataset, analyzed thematically following Terry et al. (2017), see Table 2. NVivo software supported rigorous and transparent coding. The process began with familiarization, where transcripts were reviewed to identify initial patterns and recurrent themes. In the open coding phase, relevant segments concerning career mobility, job stability, adaptation, and management were labeled. This inductive process allowed codes to emerge directly from the data, accurately reflecting participant experiences.

OPEN IN VIEWER

Table 2.

Thematic coding table of long-term content moderation work in China.

Open coding (labeling)	Open coding (categorisation)	Axial coding (conceptualization)	Selective coding (core conceptualization)
They give you at least a month to learn. Surviving the first 3 months is key.	Survival in the first 3 months	Adoption	Individual Strategies (Micro-Level)
Before 2020, the volume was much higher, but now it's decreased.	Changes in quality and quantity
Anyone still working after 6 months is considered adapted.	Adapted
Shifts may seem irregular, but you get the schedule a month in advance, so you can plan other things.
This job allowed me to switch careers and leave a declining industry—that was its biggest value to me.	Enter the Internet industry	Job Transfer
My only requirement when transferring positions was no night shifts.	No night shifts
I have to self-study many skills and even learn coding from scratch.	Improve personal skills
The income is quite good in my hometown, especially since it’s a tech company.	Well-paid	Relative Decent Work
It’s definitely better and more stable than working for a private enterprise.	Relatively better and more stable
Content moderation isn’t a great job . . . but it’s relatively decent under constrained choices.	Relatively good choice to them
It was actually the lowest-paid and most demanding job at first. But over time, I kept moving up from moderation to operations, and eventually transitioned to the AI trainer.	From moderation to operations	Multi-Level Division of Job Positions	Organizational Structures (Meso-Level)
I guess this was due to the unique value of my expertise in financial matters.	Unique expertise
When I was working in moderation, the job was quantifiable—the more you did, the more you earned. But after transferring to the risk control department, there’s only a baseline performance standard.	Change in evaluation criteria
I was thinking about quitting at the time, but then they promoted me and gave me a raise, so I decided to stay.	Promotion
I’m on an open-ended contract—the company can’t just fire me.	An open-ended contract	Forms of Employment
We sign contracts with the company’s subsidiaries—although the benefits are the same, we often joke that we’re just “insourced” employees—formally employed, but with no decision-making power, only the responsibility to execute tasks.	Insourcing staff
As far as I know, our company is the only internet company where content moderators are formal employees, and you can even find the exact number of active moderators in the company’s financial reports.	Formal employees
I don't mind that it's an outsourced job—I'm actually earning more than I did as a formal employee before.	Outsourced job
At the time, they chose Tianjin because it was cheaper than Beijing. Now that there are even more cost-effective locations, they’ve naturally moved there.	More cost-effective locations	Geographic Relocation
The management model used in Tianjin can be directly replicated elsewhere.	Replicable management model
The salaries offered by these internet companies, even for content moderators, are significantly higher than those provided by most local firms. While they certainly don’t match the pay levels in cities like Beijing, they are already attractive enough for many young people in the area.	Attractive job for many young people
That year, we were constantly handling COVID-related content. On one side were strict policy boundaries; on the other, strong public emotions—there was very little room for error.	Very little room for error	Everyday Politics	State Governance (Macro-Level)
They didn’t mean to make mistakes, maybe they were just tired or unaware, but still lost their jobs. We’re under constant political pressure.	Constant political pressure
I think it could be risky if a lot of moderators eventually shift to other industries—after all, they’ve been exposed to information that ordinary people are not supposed to see.	Risky
From our country's standpoint, content moderation is an industry that is bound to exist, and in fact, it exists around the world. However, in our case, moderation tends to be more targeted and enforced with stricter standards.	Common occurrence	Continuous Demand on Content Governance
We are just like the human firewall—most of us come and go. But as long as this wall exists, there will always be a need for moderators to serve as the filter.	GFW by content moderator
Although AI has replaced many content moderators, there are still things it cannot do. The number of moderators may decrease, but they won’t disappear entirely.	Cannot be fully replaced by AI
Those regular trainings and sudden briefings during emergencies are just ways to keep us on edge all the time.	Regular training	Professionalization and Institutionalization
Ultimately, you still need to establish your own irreplaceability, such as expertise in a particular field, entertainment, finance, international affairs, or public health.	Irreplaceability in a field
These trainings don’t enhance any real skills—they’re just about telling you how to follow instructions. All they expect is compliance, not critical thinking.	Be compliance

During axial coding, initial codes were grouped into broader categories to explore deeper relationships. Themes such as individual adaptation, organizational structures, technological mediation, and state governance emerged. These were critically refined for coherence and clear differentiation. Selective coding then synthesized categories into three analytical dimensions: individual strategies (micro), organizational structures (meso), and state governance (macro). This framework enabled a comprehensive understanding of how long-term moderators navigate structural constraints and opportunities to pursue stability and growth. To enhance rigor and reduce bias, a collaborative approach was used: the lead author conducted primary coding, while two other researchers independently reviewed selected transcripts and provided critical feedback to ensure analytical robustness.

Manufacturing decency: content moderators regarded as performing decent work

“Decency” was nevertheless the term most frequently invoked by interviewees when reflecting on their moderation work (#1, #3, #20). However, neither public discourse nor the extant scholarly literature commonly conceptualizes content moderation as “decent work.” According to the International Labour Organization’s 2008 Decent Work Agenda, the notion of Decent Work is anchored in four mutually reinforcing strategic pillars: Employment Creation, Rights at Work, Social Protection, and Social Dialogue. Paradoxically, content moderation exhibits structural deficiencies across all four dimensions, manifest in precarious employment, limited protection of labor rights, inadequate social protection, and the absence of meaningful channels for dialogue, thereby rendering it difficult to qualify as “Decent Work” in the ILO’s sense. This paradox indicates that, within the Chinese context, the meaning of “Decency” as articulated by moderators is not reducible to the ILO framework but instead entails a more multilayered and contextually embedded understanding.

First, advances in machine moderation technology have reshaped both the tasks and the work environment of moderators, reducing repetitive labor and, to some extent, the overall workload handled by humans compared with 10 years ago (Jhaver et al., 2022). As one interviewee mentioned, “Before 2020, the volume was much higher; now it’s decreased significantly” (#12). Depending on the nature and complexity of the content, moderators are typically required to process between 100 and 500 items per hour (#1, #3, #5, #8, #9, #12, #13, #14). While some jokingly describe themselves as “spacebar-hitting robots” (#2) or “low-end tools” (#5), many long-term moderators do not view their work as purely repetitive or tedious (#11, #12). Instead, they quickly internalize the rules and focus on task completion, often deriving a sense of fulfillment or enjoyment from the process (#12, #15). The initial phase of learning the rules represents a critical period for determining their suitability for long-term engagement: “The company will give you at least a month to learn; surviving the first three months is key” (#12). “During the initial adaptation period, the workload is only 80% of a full assignment” (#3). “If you didn’t achieve good performance at the start, it’s actually hard later on” (#3). “Those who are still working after six months are generally considered adapted” (#4). “Out of the 50+ people in my group, only 1–2 failed. After all, most people want to earn a salary through work, and these rules aren’t hard to understand” (#15). As one moderator summarized, “The rules are stable; the longer you work, the more skilled and comfortable you become” (#8).

However, long-term moderators often face a more persistent challenge: the physical toll and health risks associated with night shift work. Surprisingly, many long-term moderators actively choose shift-based work. As they explain, “Shifts may seem irregular, but you get the schedule a month in advance, so you can plan other things” (#2, #4). Still, not everyone adapts well: “Night shifts disrupted my hormones, so I tried to transfer out as soon as I could” (#1). For some, this flexibility supports family life: “The schedule allows married colleagues to spend more time with their children during the day” (#3). Perhaps most importantly, shift-based work affords clear boundaries: “The good thing is that once you’re off work, you don’t have to think about it at all. Your time is entirely your own” (#1, #2).

Second, decency is not limited to the working conditions and occupational safety and health, but also extends to the respect and identity that workers attain within the social structure. As one moderator put it, “Content moderation isn’t a great job, but compared to what else is available to me, it’s much better pay, better benefits. It’s relatively decent under constrained choices” (#7, #15). “After all, it’s a well-known tech company; it sounds decent when you tell others” (#1, #5). What's more, “I’m on an open-ended contract and the company can’t just fire me” (#1, #9). “The income is quite good in my hometown, especially since it’s a tech company” (#1, #3, #7). “It’s definitely better and more stable than working for a private enterprise” (#12). These narratives suggest that decency is not derived solely from social protection such as wages, contracts, and benefits, but also from the symbolic meanings attached to this job. Within local labor markets, securing a position in a major Internet company itself signifies an elevation of social status and the legitimation of one’s occupational identity. In this sense, decency operates less as an absolute measure of occupational superiority than as a relative advantage produced through social comparison. This form of “decency” further reinforces moderators’ sense of recognition, contributing to their commitment and long-term attachment to the profession.

Third, several interviewees emphasized that this work provides them with a sense of “privilege” in China. They are exposed to an informational environment that ordinary people cannot discuss or even access. This exceptional work experience generates a form of cognitive advantage compared to others within the same social stratum: as moderators, they are able to witness the unspeakable dimensions of social reality, particularly those related to political security in China.

“As friends sometimes complain about why their accounts were banned or silenced, we actually know exactly why, somehow, it was done by us” (#6). “We can see the entire trajectory of a public opinion incident on the backend, and we even need to study it carefully in order to make better moderation decisions. That’s something others have no chance to understand” (#14). For them, content moderation is not merely a livelihood but also a window through which they reinterpret both society and their own positionality. This distinct form of cognition grants them a sense of “privilege,” allowing them to derive meaning beyond economic compensation from their moderation tasks. In addition, whereas others can only accept the outcomes of content moderation, moderators possess a small, operational power granted by the system, enabling them to determine how content is handled; this limited power also constitutes a covert source of their sense of “privilege.”

At the same time, this “privilege” also constitutes a protective moat for the profession, reinforcing the stability of their employment. For reasons of risk control, “both the state and enterprises are reluctant to see a large number of trained individuals with deep knowledge of censorship logic flow freely across the labor market” (#3). Thus, moderators’ work is simultaneously a form of personal cognitive emancipation and a tightly constrained institutional arrangement. In other words, the perception of “Decency” emerges not solely from economic rewards or corporate reputation, but from the close entanglement of their labor with national Internet security, which imbues their work with additional legitimacy and a sense of mission.

Manufacturing pipeline: content moderation as a system of stratified roles and diverse divisions of labor

Content moderators in China do not represent a homogeneous group or a single, uniform job category. Instead, they are situated within a stratified and increasingly specialized labor system, one that reveals how “stability” is actively produced through differentiated roles, internal mobility, and institutionalized career pathways. Within platforms, moderation roles are clearly differentiated, assigning moderators to distinct positions along the information flow. Typical roles include frontline moderator, senior moderator, quality inspector, and the role of risk control. In addition to these moderation functions, there are supporting roles that sustain the system, such as trainer, supervisor, and data labeler. Some moderators also transition out of moderation roles into upstream positions such as data analyst, content safety policy specialist or operations role, while still remaining for the broader content governance.

Figure 1 illustrates a complex content-moderation system that operates in coordination with automated processes, which demonstrates how a message, comment, video, or image flows through the moderation system after being published. By drawing on the career trajectories and workplace experiences shared by long-term content moderators, we reconstructed the internal division of labor within the content moderation system at Chinese Internet companies. It is important to note that this does not represent the organizational structure of any single company. Rather, it is a generalized division of labor synthesized from the accounts of 21 experienced moderators. After the initial version was produced, the diagram was further refined based on feedback from these moderators.

OPEN IN VIEWER

Figure 1.

Commercial content moderation system in China.

When content enters the platform, it is first screened by an automated moderation system, which labels most items as either “Clean” or “Should Be Removed.” Content that the system cannot classify will be passed to labelers. To minimize misclassification, the platform also employs risk control moderators and appeals moderators, who handle internal secondary reviews and user feedback, respectively.

Only a small fraction of items flagged as “Unclear” by the system proceed to the manual moderation process. At this stage, tasks are routed to two different content pools: safety-oriented and quality-oriented moderation. However, all content must pass through the safety-oriented content pool, but not all content proceeds to the quality-oriented content pool. Frontline content moderators constitute the core workforce of the moderation team and handle the largest volume of tasks under the coordination of team leads. Their decisions are spot-checked by quality inspectors, often promoted from the front line, who do not re-score content or make any decisions but serve as an internal quality-control mechanism.

After content has undergone safety-oriented moderation, it may either be released directly or further routed to the quality-oriented content pool. It evaluates originality, clarity, copyright compliance, and other dimensions. Unlike safety moderation, it does not require real-time responses and follows a more stable daily routine. As a result, frontline content moderators for this pool are generally not required to work overtime or night shifts.

In addition, senior moderators occupy an intermediate layer between frontline moderators and content operations, yet little attention has been paid to. On one hand, they handle complex and difficult review tasks escalated from frontline content moderators, conducting collective assessments. On the other hand, they liaise with content operations, responsible for communicating and training frontline content moderators on platform policies, urgent directives, and updates in moderation practices. Typically promoted from frontline content moderators, they usually possess specialized knowledge in certain domains such as finance, law, healthcare, entertainment, or social affairs, enabling them to independently evaluate difficult cases within their expertise. Their work is no longer measured by piece-rate standards and cannot be easily quantified.

This system also reveals what it means to “succeed” as a content moderator in China. Survival in the competitive environment is not only a matter of efficiency under piece-rate metrics but also the ability to align with platform expectations and state-driven security concerns. Those who demonstrate both technical competence and a readiness to frame their work as a form of “public duty” are more likely to advance into senior or operational roles.

It recalls Braverman’s “internal labor market,” where undifferentiated human labor defined Fordist firms during 1970s U.S. monopoly capitalism, echoing similar yet different patterns in today’s contemporary content moderation. Content moderation in the AI era presents clear differences across positions such as wages, shift schedules, and cognitive demands vary significantly. This points to a stratified, diversified system of human-machine collaboration in moderation work. Such structural variation not only deepens our understanding of digital labor but also opens new paths for career development. As one moderator noted: “When I first joined, I had a plan. The company offers both horizontal and vertical paths, advanced moderation and quality checks for lateral growth; team leadership or training roles for vertical progression” (#10). Another shared a longer-term vision: “Content moderation is my only route out of a sunset industry and into tech where I have chance to get promotion from time to time if I did it well” (#15).

In terms of forms of employment, China’s content moderation industry typically adopts a parallel model of “insourcing” and “outsourcing,” with relatively limited use of the “crowdsourcing” model that is widely employed in Western countries due to security issues. Under the “insourcing” mechanism, moderators sign labor contracts with subsidiaries affiliated with the Internet platform, enjoying relatively stable labor relations and benefits comparable to those of formal employees. In contrast, under the “outsourcing” mechanism, workers establish labor relations with third-party human resource service companies; although the boundaries are more ambiguous, this arrangement still constitutes formal employment in form. We further examined the employment conditions of content moderators under three different modes of work organization: formal employees, insourcing, and outsourcing—(1) the formal employee model, represented by Bilibili, moderators sign contracts directly with the company. By the end of 2024, Bilibili employed 2247 moderators—28% of its workforce—making it the second-largest staff group (Bilibili, 2025). (2) The insourcing model, represented by ByteDance, involves contracts with wholly owned subsidiaries. In 2022, ByteDance had over 20,000 moderators, about 20% of its workforce (cited from Zhai, 2023). Furthermore, based on social insurance data from 21 subsidiaries, the estimated total ranges from 30,000 to 40,000 in 2024, accounting for at least 20% of the total workforce. Some interviewees suggested that major moderation bases may employ up to 50,000. (3) The outsourcing model, used by Renrui Human Resources, serves multiple platforms. As of December 31, 2024, Renrui employed 41,868 flexible workers for moderation tasks, with an average monthly cost of 10,300 RMB (~1431 USD) and a turnover rate of 6.8%, down from 8.3% in 2023 (Renrui Human Resources, 2025). The multi-level internal and external division of labor in content moderation within Chinese platform companies jointly contributes to the complexity and diversity of the content moderation labor system.

Manufacturing security: the logic of everyday politics and content governance

The 2018 shutdown of ByteDance’s “Neihan Duanzi” ² marked a critical turning point for the development of content moderation in China (#9, #15). In response to widespread public and regulatory scrutiny, ByteDance began a systematic expansion of its content moderation workforce, launching large-scale recruitment of moderators and establishing specialized moderation centers across China (Palmer, 2022). This trajectory became particularly pronounced during the COVID-19 pandemic. One moderator explained, “These years, we were constantly handling COVID-related content. On one side were strict policy boundaries; on the other, strong public emotions—there was very little room for error” (#4). Content moderation roles, with their ongoing demand, relatively stable wages, and accessible entry requirements, became a form of “safe harbor” for many jobseekers. “Many of my friends couldn’t find work at the time. I had a job and a steady income—so I considered myself lucky” (#11). Thus, content moderation jobs not only reflect platforms’ increasing reliance on risk management but also reveal their role as a form of social cushioning during times of economic disruption.

Moreover, the combined forces of state governance demands and platform compliance requirements have propelled the professionalization and institutionalization of content moderation (Wang and Zhou, 2021), transforming it from a temporary form of labor into a relatively stable, long-term occupation. Against the backdrop of ongoing campaigns, the “Clean Cyberspace” initiative (qinglang xingdong) ³ , the state’s approach to regulating the digital information environment has become increasingly structured and assertive under the campaign-style governance (Yu and Shang, 2023). “From our country's standpoint, content moderation is an industry that is bound to exist, and in fact, it exists around the world. However, in our case, moderation tends to be more targeted and enforced with stricter standards” (#12).

The institutional trajectory of content moderation builds upon earlier state efforts, such as the Great Firewall project initiated in the early 2000s, which institutionalized national-scale information filtering. “We are just like the human firewall, most of us come and go. But as long as this wall exists, there will always be a need for moderators to serve as the filter” (#7, #21). Notably, the Cybersecurity Law of the People’s Republic of China, implemented in 2017, codified the responsibility of network operators to manage information content, respond to emergencies, and ensure the security of online spaces (Cyberspace Administration of China, 2016). More recently, the Provisions on the Administration of Algorithmic Recommendation for Internet Information Services (Cyberspace Administration of China, 2022) mandated that platforms establish content review mechanisms to prevent algorithmic amplification of material deemed to “endanger national security” or “disrupt social order.” These legal developments have provided a formal policy foundation for the institutionalization and professionalization of content moderation labor.

The long-term institutionalization of state governance logics and concerns over ideological security constitutes a fundamental driver behind the professionalization of content moderation in China. Around 2018, a so-called “formalization campaign” marked a pivotal moment, not merely standardizing workflows, but embedding ideological security into everyday moderation routines. Leading Internet companies such as ByteDance, Tencent, and Baidu began establishing large-scale content moderation centers in cities like Tianjin or Chengdu. Content moderation on these platforms thus became more than a matter of corporate compliance; it was increasingly tasked with preventing the spread of “erroneous ideological trends,” safeguarding against the perceived threat of “color revolutions,” and resisting foreign ideological infiltration. Within this ideology-oriented regulatory framework, moderators are expected not only to follow technical and procedural protocols but also to internalize political positions and align with the “mainstream ideological line” (zhuxuanlü). “Those regular trainings and sudden briefings during emergencies are just ways to keep us on edge all the time” (#13). Dismissed moderators—often contacted by the Cyberspace Administration of China—become cautionary tales. “They didn’t mean to make mistakes, maybe they were just tired or unaware, but still lost their jobs. We’re under constant political pressure” (#6).

In the context of this ongoing institutionalization, the Chinese state has progressively integrated ideological governance into routine technological and organizational operations, encouraging platforms to establish comprehensive moderation infrastructures and accountability chains (Fu et al., 2025). Pre-employment training for moderators now routinely covers not only content guidelines and operational procedures but also understanding national policies, affirming political loyalty, and developing the ability to identify “hostile” content. “I think it could be risky if a lot of moderators eventually shift to other industries, after all, they’ve been exposed to information that ordinary people are not supposed to see” (#4). Content moderators not only receive systematic training and sign confidentiality agreements, but also engage in long-term, sustained exposure to highly sensitive forms of information in their everyday work, including emerging public opinion incidents, the concrete operational boundaries of policy red lines, and coordination mechanisms between platforms and regulators. It is under this condition of structural exposure that labor mobility is constructed by platforms and the Chinese state as a potential security risk. Importantly, this notion of “risk” does not primarily refer to actual breaches of confidentiality, but rather reflects a forward-looking, precautionary governance imaginary. It reveals how content moderation is institutionally treated as a form of security infrastructure, which in turn underpins platforms’ emphasis on recruiting “politically reliable” and “organizationally stable” personnel in order to maintain organizational controllability and ideological alignment.

Under the dual pressures of increasingly codified state regulation and platforms’ own imperatives for operational stability, content moderation training and organizational development have come to represent the “last mile” in digital governance infrastructures. Our study highlights that content moderation is shaped not only by market forces but also by the regulatory and normative frameworks of the government. The Chinese government influences moderation indirectly through policies, laws, and internal guidance tools such as “blacklists and whitelists” for platform reference, while also shaping moderators’ ideological orientations through more subtle mechanisms. Accordingly, the institutionalization of content moderation should not be understood merely as a market-based response to information risk or cost efficiency, but as a politically driven governance practice, in which state directives are deeply embedded within platform operations.

Conclusion: possible but improbable stability

This study set out to explain a paradox that unsettles dominant accounts of platform labor. Why and how has commercial content moderation in China, a form of work widely understood as precarious, invisible, and emotionally taxing, become for many workers a relatively stable and even desirable long-term occupation. Based on ethnographic fieldwork and interviews with 21 moderators across different roles and institutional settings, we show that this stability is best understood as possible but improbable. It is possible because it is actively produced through institutional arrangements. It is improbable because it emerges from a work arrangement that is structurally organized around disposability, routinized exposure, and intensifying control. The stability observed here is therefore neither spontaneous nor emancipatory. It is manufactured as a structural condition through the coupling of individual survival strategies, organizational arrangements, and state governance imperatives. Furthermore, under conditions of heightened geopolitical rivalry, proliferating sites of armed conflict, and successive waves of economic and technological shock, stability and precarity should be treated as historically contingent and politically mediated classifications of work. For workers in many non-Western contexts, what appears as stability often takes the form of an institutionally managed continuity rather than a legally grounded entitlement.

Rather than treating stability as a linear progression from precarity to formality, our findings suggest that stability itself is a governed and institutionalized outcome. While existing scholarship typically conceptualizes precarity as instability, insecurity, and disposability (Kalleberg, 2009; Standing, 2011), we argue that stability can emerge not as the resolution of precarity but as its strategic recalibration within a governance infrastructure. Moderators do not escape precarity. They learn to inhabit it in ways that render their work continuous, legible, and administratively manageable. Similar to Ravenelle’s (2019) notion of hustle and Anwar and Graham’s (2020) concept of tactical endurance, Chinese content moderators actively manage constraints within a tightly regulated system and carve out a form of continuity that is fragile yet durable. This continuity is enabled and delimited by institutional logics rather than achieved in spite of them. In this sense, possible but improbable stability names a distinctive mode of labor institutionalization. Governance can stabilize an otherwise destabilizing occupation while leaving intact the underlying asymmetries of control, risk, and replaceability that define the work in the first place.

This form of continuity, however, should not be read as the product of autonomous agency alone. It is better understood as a managed accommodation that makes long-term employment possible, even if it remains improbable in a line of work widely associated with high turnover and routine strain. In this sense, our findings echo Burawoy’s (1979) analysis of manufacturing consent, in which compliance is not secured solely through coercion but produced through situated incentives, internal differentiation, and symbolic rewards. Much like factory workers who played the game to meet production targets under monopoly capitalism, content moderators do not primarily aim to escape control. They instead develop practical strategies to survive within it, and to convert uncertainty into a limited form of predictability.

Although the work remains demanding, many interviewees describe a sense of security and satisfaction that contrasts with the volatility of other informal occupations. This sentiment is frequently articulated through the term decent work, which in the Chinese context carries both material and symbolic meanings. It refers to a steady income, access to social insurance, affiliation with reputable technology companies, and a nonmanual working environment. This usage does not align fully with the ILO’s normative criteria, particularly regarding participation, voice, and long-term protection. Yet it operates as a meaningful category that workers negotiate under structural constraints. Recent ILO reports also underscore that decent work should be assessed not only through formal guarantees, but through how workers experience dignity, voice, and predictability in local contexts (ILO, 2023). Our findings therefore resonate with Betti’s (2018) call to historicize precarious work and to situate labor norms within uneven national trajectories of industrialization, governance, and social welfare. In this light, decent work is not an endpoint but a relational construct that takes shape through accommodation among state governance, platform control, and worker adaptation.

This negotiated sense of decency cannot be separated from the broader architecture of Chinese digital governance. Content moderation in China is not merely a corporate necessity but a politically consequential responsibility. Platforms are held accountable for online order, and moderators become key actors in fulfilling the dual imperatives of market performance and ideological security. The professionalization of moderation systems, the expansion of operations into second- and third-tier cities, and the diffusion of multilayered review protocols reflect managerial rationality, but they also signal the institutional embedding of state power within digital labor infrastructures. In this context, moderation is no longer an auxiliary or temporary task. It becomes a routine component of governance for managing information flows, shaping public sentiment, and sustaining political legitimacy.

At the same time, Chinese platforms have rapidly integrated automation into moderation workflows, not as a simple substitution of humans but as a continuous reconfiguration of what human work is for. Artificial intelligence systems now handle vast volumes of low risk and repetitive tasks, accelerating the speed and scope of initial screening. Human moderators remain indispensable, however, when content is ambiguous, politically sensitive, or ideologically charged, especially where regulatory expectations demand contextual judgment. Moderators also contribute to the training and refinement of these systems through data labeling, escalation protocols, and quality control. Their judgment is folded into ongoing feedback loops that sustain platform governance. We argue that this integration of automation reshapes moderation from high volume execution toward judgment aligned with governance priorities. Rather than increasing visibility or recognition, this shift embeds human labor more deeply into the infrastructural logic of content governance, thereby reinforcing the conditions under which possible but improbable stability can be manufactured.

This study advances a non-Western account of platform labor that unsettles familiar oppositions between precarity and formality, automation and obsolescence, and corporate management and state power. The stability observed here is manufactured on top of, and continually exposed to, deeper layers of instability. As tech companies accelerate automation and reorganize workflows, the boundaries of human indispensability can contract, and tasks may be redefined, relocated, deskilled, or eliminated with little warning. At the same time, shifts in regulatory priorities, security-oriented governance, competitive pressure, and the wider reconfiguration of global technology supply chains can abruptly alter what kinds of labor are valued, protected, or rendered redundant. In this sense, manufactured stability is best understood as conditional and reversible. It makes continuity possible while leaving intact the very conditions under which continuity can be withdrawn. Amid the global turbulence and uncertainty, stability and precarity are increasingly unstable categories themselves, especially in non-Western societies. Attending this terrain requires that we ask not only how precarious/stable work is produced, but how stability is selectively manufactured, for whom, and at what social and ethical cost. If the horizon of work is being rewritten through both automation and security, then a more urgent task for scholarship is to render visible the infrastructures that govern endurance, to take seriously workers’ dignity and voice within them, and to imagine forms of security that do not depend on the permanent normalization of control.