At 7:49 pm on 24 November 2022, a fire in a residential building in Urumqi, the capital of the Xinjiang Autonomous Region, took the lives of ten individuals and injured nine others. Social media users alleged that fire lanes had been blocked by barriers and that exits had been sealed, preventing residents from fleeing the building. The incident prompted widespread criticism of China’s zero-COVID policy, which had imposed extended lockdowns across the country for nearly three years.Footnote 1 Besides dissatisfaction with these measures, the disaster catalysed broader dissent. By 26 November, thousands of people across China had taken to the streets in what became the geographically most expansive mass protests since the Tian’anmen 天安门 uprising in 1989.
In the time between the fire and the ensuing protests, Sina Weibo, China’s second-largest social media platform, became a hotbed of digital dissent. Hashtags related to the fire dominated Weibo’s “Hot Search” (Weibo resou 微博热搜), accumulating tens of millions of views.Footnote 2 Within just 52 hours, nearly 390,000 messages, most of them very critical, had flooded the platform. Activity peaked on the morning of 26 November, with over 800 messages per minute. At around 3:00am on 27 November, Weibo imposed a blanket ban on the topic.
Considering the authorities’ tight control over online spaces in China, both the scale and intensity of this dissent are striking.Footnote 3 Most social media accounts have very few followers, significantly reducing their visibility, while “Big V” opinion leaders are kept under close surveillance.Footnote 4 As the state has systematically dismantled the high-visibility nodes through which digital dissent once spread, how were users mobilized after news of the disaster became public? How could such extensive participation achieve visibility without coordination? And, given the absence of opinion leaders who might orchestrate the discourse, where did citizens direct their outrage?
Using a dataset of 274,721 Weibo entries (of an estimated 390,000) generated between 24 November and 27 November 2022, this study combines qualitative and quantitative content analysis to provide answers to these questions. It introduces the concept of “subversive amplification,” the process by which ordinary users hijack state-sponsored content to achieve mass visibility for dissenting messages.
The study finds that subversive amplification was triggered not through the now tightly controlled high-follower opinion leaders or other individuals, but through official communication failures that made citizens angry. The associated hashtags provided the anchors to which dissent could be attached. Ordinary users converged independently on these anchors, hijacking state-promoted hashtags en masse and exploiting their visibility to achieve reach without needing to rely on traditional amplifiers. However, because subversive amplification bypasses opinion leaders, there are inherent limitations to its effectiveness: without actors to coordinate individual voices towards collective messaging, the contention manifested as a fragmented menu of complaints unable to condense into a coherent prognostic frame.
The article proceeds by situating the study within research on digital contention in China and then introduces the dataset and methods. The empirical section traces how users hijacked and amplified official content, highlighting in the process the official communication failures that enabled this mobilization. It then measures the prevalence of critical content across account types and maps that content across four categories: emotional disaffection, systemic and ideological critique, contestation of information control and policy critique. The concluding section discusses implications for digital resistance on heavily controlled digital platforms.
Digital Control and Resistance in China
Scholars broadly agree that intensified state control has made digital dissent in China extremely difficult. Through censorship, content flooding and the targeting of influential accounts, authorities have imposed strict limits on expressions of systemic criticism and civic discourse online.Footnote 5
This represents a marked change from the early internet era, when scholarship documented the shared advantages of online public spaces for both citizens and the government.Footnote 6 Yet, even then, online public spaces were heavily contested,Footnote 7 and dissent could be explained as much by strategic tolerance as by the government’s inability to keep pace with society’s rapid adoption of new digital technologies.Footnote 8
Research has also illustrated how netizens have adapted, turning to parody, satireFootnote 9 and indirect forms of critique, such as praising past leaders to implicitly criticize current ones.Footnote 10 Other research convincingly demonstrates that, besides serving political purposes, such tactics can foster solidarity and facilitate social interaction,Footnote 11 as well as criticize and entertain simultaneously.Footnote 12
While dissent in China has always faced obstacles, recent scholarship highlights how, in the Xi Jinping 习近平 era, the combination of increased internet penetration, platform diversification and expanding government control has made the obstacles seemingly insurmountable. As internet penetration has increased, China’s netizens have increasingly come to mirror broader Chinese society.Footnote 13 Critical content is forced to compete with celebrity gossip and other forms of entertainment,Footnote 14 while nationalist (micro-)bloggers such as ziganwu 自干五 and the “Little Pinks” (xiao fenhong 小粉红) have joined the ranks of government-recruited online commentators to challenge and harass online activists.Footnote 15 The government has also established specialized channels for “loyal dissent,”Footnote 16 reducing its dependence on social media as a source of information and, in consequence, the incentive to tolerate online criticism.Footnote 17 Finally, the authorities have systematically combined regulatory tightening, technological upgrading, content flooding and the strategic targeting of information gatekeepers to promote so-called “positive energy” (zheng nengliang 正能量).Footnote 18
On the regulatory front, the authorities have pursued a multi-pronged strategy of real-name registration,Footnote 19 stricter anti-rumour regulationsFootnote 20 and heightened monitoring requirements for internet-service providers and platforms. Circumvention tools have been purged from app stores and a China-specific network standard (IPV6+) implemented, enabling authorities to trace information to its source.Footnote 21
Simultaneously, state-affiliated accounts have colonized online public spaces, with major propaganda channels attracting over 100 million followers each on Sina Weibo.Footnote 22 State media have been highly successful in capturing attention and generating engagement,Footnote 23 and government narratives have largely displaced critical discussions of Chinese politics on Sina Weibo.Footnote 24 Research on diffusion cascades finds that viral content depends on a small number of high-follower accounts acting as broadcast hubs,Footnote 25 a vulnerability the government has exploited by systematically purging or co-opting these gatekeepers.Footnote 26 As a result, outspoken public intellectuals have retreated into tightly connected issue networks operating below the radar of censorship.Footnote 27
However, the government’s control over information flows is not complete. The system is designed to raise the cost of accessing sensitive content rather than to block it entirely.Footnote 28 Heightened traffic during peak periods can temporarily overwhelm the censorship infrastructure,Footnote 29 and attempts to enlist netizens in co-producing propaganda can backfire, provoking criticism or sarcasm.Footnote 30
The online mass event (OME) is a form of dissent that exploits these vulnerabilities.Footnote 31 The term refers to episodic bursts of contention triggered by scandals or abuses of power, whose impact depends on the rapid self-mobilization of a large number of ordinary users.Footnote 32 Early examples include public outrage over migrant worker Sun Zhigang’s 孙志刚 death at the hands of urban management officers, the “My Father Is Li Gang” (wo ba shi Li Gang 我爸是李刚) scandal and the 2011 Wenzhou 温州 train collision, all of which triggered large-scale online mobilization.Footnote 33 More recent instances include the 2015 Tianjin explosionFootnote 34 and the 2018 Shouguang City 寿光市 flood.Footnote 35
The nature of these events raises questions that existing research cannot fully answer. First, if traditional amplifiers no longer influence which events gain traction, what determines whether an incident escalates into an OME? The Guizhou bus crash two months prior to the fire, which also occurred in connection with pandemic policies and resulted in more fatalities than the fire, did not produce an OME.Footnote 36 What made the Urumqi fire special? While this study cannot explain why widespread contention emerged in one case and not the other, it can identify the conditions that enabled it in the Urumqi case.
Second, online dissent once relied on public intellectuals as intermediaries: content produced by ordinary users was amplified when public intellectuals picked it up and shared it with their large follower base.Footnote 37 Beginning in the early 2010s, the authorities largely inactivated these nodes by silencing or co-opting the broadcasters. Yet visibility requires convergence on shared focal points.Footnote 38 How can online dissent achieve scale when traditional amplifiers do not provide coordination?
Third, in most forms of contentious collective action, mobilization and framing tend to be intertwined: movements coalesce around shared diagnoses and visions of what should be done.Footnote 39 In online mass events, by contrast, (self-)mobilization occurs through outrage, but what to mobilize for (or against) does not naturally follow from the trigger event. Without any guidance from opinion leaders, where do citizens direct their energy?
A key factor in understanding how subversive content can gain virality on Weibo under these constraints is that it is driven by hashtags and engagement metrics, irrespective of content. The same mechanisms that amplify state narratives can, in principle, also amplify dissent. This does not make the rules of engagement very different from other platforms, such as Twitter or Facebook. Scholarship on “connective action”Footnote 40 and “hashtag activism”Footnote 41 on those platforms has examined how influential accounts use hashtags to mobilize participation, coordinate visibility and supply prognostic frames.
However, China’s tightly controlled social media platforms largely preclude such forms of activism.Footnote 42 Hashtags on Weibo are subject to censorship, and influential user accounts are under close surveillance. State accounts dominate the discourse through a combination of high follower numbers and algorithmic prioritization of their hashtags on Weibo’s Hot Search list, which showcases the most popular hashtags along a user’s timeline.Footnote 43 Accounts with few followers have an infinitesimally small chance of gaining widespread attention. If influencers do not help to bring a topic to public attention, and if a social media company does not feature it on users’ timelines, it has virtually no chance of going viral.Footnote 44
This study argues that users exploit these features through what it terms “subversive amplification,” a process by which they appropriate state-sponsored content to turn propaganda against itself. Instead of creating their own hashtags, users hijack hashtags from state-promoted accounts, either by quote-posting or inserting them into their own critical messages. By exploiting the algorithmic visibility of official accounts, they collectively achieve reach. The resulting OME is not driven by influencers who serve as aggregators and amplifiers; instead, users are directly activated by state narratives that trigger latent grievances. However, since this mechanism bypasses opinion leaders who usually provide collective action frames, the resulting dissent manifests as a menu of grievances but does not converge on shared demands. Prior studies have examined how users engage with official content but not how they strategically use it for subversive purposes.Footnote 45
Data and Methods
To contribute insights into how online mass events emerge, what triggers them and what citizens express through them, this study analyses Weibo discourses following the 24 November 2022 Urumqi residential building fire. The case is instructive because while scandals, accidents and disasters often stimulate online debates, they seldom escalate into an OME.
Data
The dataset comprises 274,721 Weibo entries (65,254 posts and 209,467 shares) published between 24 November and 27 November 2022.Footnote 46 Entries were identified through searching for the keywords Wulumuqi 乌鲁木齐 and Xinjiang 新疆. Real-time collection began on 26 November at 2:25pm, Beijing time, at five-minute intervals to minimize censorship bias. Posts from 24 November to 26 November were collected retrospectively. This approach may have missed posts that were using oblique references or coded language to evade keyword-based censorship; however, such posts would have lacked hashtag visibility and would have unlikely achieved significant reach. As the study focuses on subversive amplification, which by definition operates through identifiable hashtags, the exclusion of any such posts is unlikely to change the results of this study.
Using its metadata, each message was labelled for account type (official account, opinion leader, non-verified account), hashtag origin (the type of author that initiated the hashtag) and message type (original post or share). Furthermore, each entry contains the ID of its “parent” post, allowing the tracing of the trajectory of specific messages and the identification of quote-post hijacking.
Shares were collected retrospectively in August 2023. These capture the full cascade of downstream sharing (shares of shares), but they only contain content that survived censorship. To assess the extent of censorship, the status of real-time collected posts was checked four months after the initial collection, and deletion rates were calculated for each hour of the dataset.
Figure 1 illustrates the hourly post volume and deletion rates during the observation period. The deletion rate fluctuates consistently between 50 and 65 per cent from the onset of reporting about the fire to the topic ban. This consistency suggests that retroactive collection on 26 November was not affected by censorship bias. A difference-in-differences analysis comparing sensitive and non-sensitive content across retroactively and real-time collected posts confirms this (see Appendix 1).
Hourly Volume and Deletion Rate of Posts Over Time
Figure 1 Long description
Publication Time on the horizontal axis, ranging from 11-24 0:00 to 11-28 0:00. Number of posts on the left vertical axis, ranging from 0 to 4000. Deletion rate percent on the right vertical axis, ranging from 0 to 70. Legend entries: not deleted; deleted; Deletion rate. A vertical dashed line is labeled FIRE INCIDENT. A stacked area chart shows two stacked categories labeled not deleted and deleted. The stacked total stays near 0 for much of the timeline, then rises and falls in several clusters. A smaller cluster appears around 11-25 12:00, with a peak below 2000 posts. Additional smaller peaks appear between 11-25 12:00 and 11-26 12:00. The largest cluster occurs between 11-26 12:00 and 11-27 12:00, with the stacked total reaching close to 4000 posts. After 11-27 12:00, the stacked total drops back toward 0, with small fluctuations through 11-28 0:00. A line labeled Deletion rate runs across the plot using the right vertical axis. It fluctuates across the full time range, with values mostly between 20 and 70 percent. It rises to near 70 percent around 11-26 12:00, then drops to around 30 to 40 percent after 11-27 12:00 and continues fluctuating through 11-28 0:00.
The true number of shares for each hour in the observation period was estimated by applying the deletion rate to the number of shares retroactively collected in August 2023. To validate this approach, the estimated volume of total reposts (collected plus estimated deleted) during the recollection period was compared against available metadata. The difference between the estimate and the number of shares indicated by adding up share metadata from collected messages was just 6.1 per cent, confirming the validity of this approach. The test was limited to data retroactively collected on 26 November, because these messages were relatively unaffected by censorship, while most sharing had already occurred, meaning the post sample was largely complete and the embedded share counts reliable. Deletion rates for different account types were also calculated. Posts by non-verified accounts were subsequently deleted at a higher rate (57 per cent) than those by opinion leaders (27 per cent), either through platform removal or self-deletion.Footnote 47
Methods
The study employs a mixed-methods approach to analyse discourse dynamics across the observation period. To identify the types of critical content users attached to hijacked hashtags and quote-posts, a random sample of 400 original, hashtag-bearing posts was subjected to qualitative content analysis. The sample size was calculated to ensure statistical representativeness at a 95 per cent confidence level. Through an inductive coding process, four overarching categories of critical content were identified: emotional disaffection, systemic critique, contestation of information control and policy critique. The two largest categories, emotional disaffection and systemic critique, each account for approximately one-third of critical content. Contestation of information control and policy critique, the remaining third, are roughly equal in size.
Based on these insights, a classification prompt for a generative large language model (LLM) was constructed. The full dataset of 274,721 entries was then classified using the Gemini-3 Flash API. Correctly classifying subtypes of critical engagement proved challenging because of the difficulty of detecting subtle irony and historical allusions. The model was, therefore, prompted to categorize messages into three broad stances: critical of the government or its policies, supportive, or neutral/ambiguous (see Appendix 2 for the full prompt and coding schema). Using two hand-coded samples of posts, the validity of the machine-coded results was then tested through a two-stage verification process (see Appendix 3). Validation confirms that while the model is effective at identifying dissent, it sometimes misclassifies sarcastic or veiled criticism as support. As described in Appendix 3, a statistical correction was applied to aggregated counts to gain a more accurate representation of the three stances in the overall discourse. Despite these limitations, generative LLMs remain superior to traditional sentiment analysis tools for this task, not least because sarcasm masks criticism in superficially positive sentiment.
Subversive Amplification After the Urumqi Fire
The following sections trace how subversive amplification unfolded after the Urumqi fire. They investigate how users converged on state-sponsored content, why they did so, and where they directed their criticism without coordinated framing by opinion leaders.
Rapid convergence on state-sponsored content
In the Xi era, state media have come to dominate social media platforms. In the domain of political discourse, their visibility surpasses that of opinion leaders. While the latter frequently challenged government narratives in the pre-Xi era, they now tend to amplify them.Footnote 48 Ordinary users tend to be consumers and amplifiers of such discourse.
The popularity of hashtags generated by different account types indicates their relevance for public discourse on Weibo. Before the fire, Xinjiang-related hashtags by official accounts and opinion leaders were roughly equally popular, and the majority of posts bore no hashtag at all. After the fire, the popularity of state-sponsored hashtags surged. Figure 2 illustrates the marked change in discourse patterns: both the volume of posts and the rate of sharing increased by orders of magnitude. Simultaneously, engagement swiftly concentrated on state-generated hashtags. This rapid convergence on state-sponsored content is an important feature of subversive amplification: focal points are needed for dispersed criticism to gain visibility, and state-sponsored hashtags provided exactly that.
Temporal Dynamics of Hashtag and Quote-Post Hijacking After the Urumqi Fire
Figure 2 Long description
Two line graphs depict trends in hashtag and quote-post hijacking from November 24 to November 29. The left graph shows hashtag hijacking with the x-axis labeled with dates and times, the primary y-axis labeled as percentage from 0, and the secondary y-axis labelled Absolute Volume from 0 to 4000. The right graph shows quote-post hijacking with the and the primary y-axis labeled similarly and the secondary y-axis labeled as number of posts from 0 to 35000. Both graphs feature lines representing ′Official′, ′Verified′, ′Non Verified′ and ′No Hashtag′ categories. Key trends include peaks in percentage and post numbers around November 26. The graphs illustrate a correlation between the two hijacking types, with similar patterns of rise and fall. The legend differentiates the lines by category.
Figure 2 shows that while hashtag hijacking fluctuated in line with official updates, quote-post hijacking was both more sustained and reached volumes that were significantly greater than hashtag hijacking.Footnote 49 During the peak phase of the OME, over 80 per cent of all shared content originated from official accounts, and the vast majority of engagement with these posts (84.1 per cent) took the form of quote-posts rather than simple reposts. Between these peaks, discourse temporarily reverted to non-contentious topics such as tourism or Xinjiang cuisine. Yet the continuous release of provocative updates by state actors sustained contention for more than two days despite the short half-life of social media virality. Figure 2 also illustrates the changing role of opinion leaders: typically drivers of non-contentious discourse, they largely withdrew from agenda setting as dissent intensified. Although they remained active, they did not create any viral hashtags.
Figure 3 maps the cumulative distribution of hashtags in original posts against their subsequent shares, further illustrating this concentration of discourse on state-sponsored hashtags. It also provides information on the issues that triggered escalation (see Appendix 4 for more information on the origins and engagement statistics of these hashtags, as well as their English translations). The top three hashtags alone account for nearly 70 per cent of shares, despite representing only 60 per cent of posts. By far the most popular hashtag referenced the 26 November declaration that Urumqi was “basically virus-free.” Users’ sharp reaction to this sudden policy change after more than three months of strict isolation can be explained by their feelings of having been betrayed (more on this below). Combined, the three hashtags relating to the press conference one day earlier commanded almost as many shares. At that event, fire chief Li Wensheng 李文胜 attributed the deaths to residents’ “low self-preservation ability,” triggering outrage at what was widely interpreted as victim blaming. Around 70 per cent of all messages focused on these two proclamations, while the hashtag reporting the fire drew far less engagement. This concentration suggests that users’ ire was directed not at the tragedy itself but at the official messaging surrounding it. Although two non-official hashtags appeared in a substantial share of original posts (highlighted in blue), they failed to attract engagement.
Cumulative Distribution of Hashtags and Subsequent Shares
Figure 3 Long description
Two-line graph with a legend listing “Cumulative Posts” and “Cumulative Shares”. The x-axis shows a sequence of hashtag labels. The y-axis label reads “Percentage of Total” and the scale runs from 0 to 70. Cumulative Posts values at the plotted points are approximately 28, 47, 60, 65, 69, 70, 78, 80, 81, 85, 87, 88, 89, 90. Cumulative Shares values at the plotted points are approximately 40, 61, 69, 69, 71, 71, 71, 71, 71, 71, 83, 89, 92, 95.
Before examining the features that made these responses so contested, the next section addresses who drove this convergence on a small number of hashtags, and whether it depended on the usual amplifiers.
Amplification without influencers
As explained in the literature review, ordinary users depend on accounts with large audiences to achieve visibility. While the previous section investigated what content users converged on, this section examines who drove this convergence. It shows that dissent was driven by the self-mobilization of ordinary users, and that although opinion leaders did not produce popular hashtags, they did participate in subversive amplification. However, the activities by these groups unfolded in parallel rather than in a hierarchical fashion: they independently targeted official content without amplifying each other.Footnote 50
Figure 4 traces how each account type amplified state-originated content. The sharing cascade reached a maximum depth of 11 levels; however, nearly all activity occurred in the first two rounds: 82.5 per cent of shares were direct hijacks of official content (Round 1) and 16.4 per cent were secondary shares (Round 2). The shallowness of the cascade confirms that subversive amplification operates through mass parallel action rather than via viral chain transmission. The follower distribution across account types (Appendix 5) confirms this observation: non-verified accounts have negligible individual reach, yet their aggregate engagement generated broadly visible dissent.
Hijacking and Amplification by Account Type
Figure 4 Long description
A flow diagram organized into three sequential columns from left to right: Source, Round 1: Hijacking, and Round 2: Subsequent Amplification. The Source stage contains a box on the far left labeled Official content. In Round 1: Hijacking, arrows lead from Official content into two categories: a large box labeled Non-verified, representing 96.6 per cent (n=257,198), and a smaller box labeled Opinion leaders, representing 3.4 per cent (n=9,065). In Round 2: Subsequent Amplification, there are two destination boxes: a top box labeled Non-verified, totaling 59.8 per cent (n=25,007), and a bottom box labeled Opinion leaders, totaling 40.2% per cent(n=16,788). The connections between Round 1 and Round 2 show that 45.0 per cent flows from Round 1 Non-verified to Round 2 Non-verified, while 48.2 per cent flows from Round 1 Non-verified to Round 2 Opinion leaders. Additionally, 2.1 per cent flows from Round 1 Opinion leaders to Round 2 Non-verified, and 4.6 per cent flows from Round 1 Opinion leaders to Round 2 Opinion leaders. A footnote at the bottom reads: “Note: Counts adjusted for deletion (non-verified: 59 per cent, opinion leaders: 29 per cent). Official accounts excluded (<0.5 per cent).”
Round 2 reveals whose content was further amplified. Despite contributing little to the initial hijacking, opinion leaders captured a disproportionate share of secondary engagement. Yet Round 2 comprised only a sixth of Round 1’s volume, so amplification did not depend on them. State media, with their massive follower bases, inadvertently provided the broadcast infrastructure. Ordinary users achieved reach by attaching critical payloads to content already posted by the state. Individually, their contribution was inconsequential, but around a third of a million of them acting in parallel made dissent broadly visible. Hence, users achieved scale not by generating their own hashtags or amplifying opinion leaders but by leveraging content already made visible by the state.
Pervasive criticism
To contextualize the qualitative content analysis that follows, this section quantifies the scale of digital dissent after the fire using machine-coded labels for all posts and quote-posts in the dataset. Figure 5 reports the distribution of stances by account type and engagement type, adjusted for the coding errors of the machine-learning model.
Amplification by Type, Stance and Account Type
Figure 5 Long description
Two side-by-side bar charts displaying Adjusted Percentage (%) on the vertical axis, ranging from 0 to 100, against Account Type on the horizontal axis, which includes non-verified, verified, and official. The legend defines Stance of Content using three patterned/colored fills for Critical, Supportive, and Neutral. The left plot, titled Hashtag Hijacking, shows that for non-verified accounts, Critical content is highest at 90.3%, followed by Neutral at 6.9% and Supportive at 2.9%. For verified accounts, Critical content is 75.4%, Neutral is 19.1%, and Supportive is 5.5%. For official accounts, Neutral content is highest at 90.3%, followed by Supportive at 8.8% and Critical at 0.9%. The right plot, titled Quote-post Hijacking, shows that for non-verified accounts, Critical content is highest at 90.6%, followed by Neutral at 8.7% and Supportive at 0.7%. For verified accounts, Critical content is 89.7%, Neutral is 9.4%, and Supportive is 0.9%. For official accounts, Neutral content is highest at 96.8%, followed by Supportive at 1.8% and Critical at 1.4%.
The results confirm that engagement with official content was nearly always subversive. Non-verified users overwhelmingly expressed criticism regardless of engagement type, and support for the government was negligible, never exceeding 3 per cent. Opinion leaders showed similar patterns, although they relied on quote-posting even more heavily than ordinary users to react directly to official content. That opinion leaders engaged critically despite tight monitoring and the risk of deplatforming might be explained by a changing attitude or risk calculus as the OME unfolded. Some might have felt compelled to speak out despite the risks; others might have felt protected within the wave of dissent.
Official posts show the highest levels of neutrality across both engagement modes. The small percentage labelled as “critical” primarily reflects classification errors. The two notable exceptions were former Global Times editor Hu Xijin 胡锡进, who posted soft criticism of the length of Urumqi’s lockdown, and the scripted nature of the subsequent press conference.Footnote 51
Voicing Dissent
Having established that subversive amplification was both pervasive and overwhelmingly critical, the analysis now turns to the content of that criticism. This serves two purposes. First, it provides deeper insights into what activated users to engage in subversive amplification. Second, it addresses the question of direction: without opinion leaders to guide framing, where did citizens direct their energy?
Emotional disaffection
Approximately one-third of critical content expresses emotional disaffection with the government: outrage, disillusionment, exhaustion and despair. Unlike the systemic critique discussed below, such content conveys not any analytical critique but rather a feeling of having been let down.
Many posts express moral outrage at the callousness of official statements. As explained above, hashtags related to the press conference following the fire attracted more engagement than those reporting the fire itself. In particular, users took offence at fire chief Li Wensheng’s declaration that “some residents had a low self-preservation ability,” which they widely interpreted as victim blaming and a deflection from government failures. The absence of any conciliatory gestures likely intensified the reaction.
It appears that officials drafted their statements with their superiors in mind rather than an exhausted and grieving public. Their statements deflected blame and showcased the arrest of individuals who had shared evidence linking fire safety violations to pandemic policies. One user observed: “No bowing, no apologies – only talk of a 3-year-old child’s ‘weak self-rescue awareness’… along with ice-cold tones and statistics.”Footnote 52
The sudden lifting of the lockdown on 26 November provoked similar outrage, again because officials seemed not to have considered how their proclamation would be received by citizens who had endured more than three months of isolation at home. The fact that Urumqi was declared virus-free overnight made them question why the lockdown had not been lifted earlier. They accused the government of restricting freedom of movement not to protect lives but to exert arbitrary control, the cost of which became undeniable when ten people lost their lives in the fire: “Turns out lifting the lockdown requires a pile of corpses.”Footnote 53
Many argued that the disaster was evidence that arbitrary pandemic controls put them all at risk. One post crystallises this logic: “We weren’t on that midnight bus in Guizhou, nor were we in that burning building in Xinjiang, simply because of good luck. But we can’t rely on luck forever. In reality, we are all on that bus, and we are all in that building.”Footnote 54 Another post ascribes intentionality to this collective vulnerability: “On this train headed towards the future, neither you nor I are the passengers – we are the fuel.”Footnote 55 Such content frames catastrophe as the inevitable consequence of a state that subordinates its citizens to its opaque political priorities.
Others expressed disappointment with unfulfilled promises of upward mobility, grievances that predated the fire and even the pandemic. Some wrote that they had long been ardent nationalists but now felt betrayed. One asked whether their suffering was “karma for being a ‘Little Pink’ [nationalist] for nearly 20 years.”Footnote 56 Related references to “run-ology” (runxue 润学) convey a desire to emigrate in search of a meaningful life. Such content conveys a sustained erosion of patriotism, with pandemic policies accelerating the trend. These users implicitly challenged the state’s grand narratives by situating themselves in the “garbage time of history.”Footnote 57
Given the economic toll of the pandemic, surprisingly few messages focused on economic loss. However, the relative scarcity of such content does not mean that material grievances were absent; rather, they were perceived as subordinate to a more fundamental dissatisfaction with China’s political order.
In more abstract terms, these emotional expressions convey an erosion of moral attachment to the state. Given that morality is increasingly central to the Party’s legitimacy,Footnote 58 accusations of its perceived absence serve to highlight the chasm between official rhetoric and lived experience. The result is cynicism towards the “mandatory optimism” on which the China Dream narratives hang.Footnote 59 While these expressions convey a feeling of betrayal, the content discussed next translates that feeling into direct challenges to the party-state’s legitimacy.
Systemic critique
A roughly equal share of critical content targets the institutional and ideological foundations of China’s political order. Although such criticism piggybacks onto pandemic-related state media content, it targets the party-state at large. Among such content, accusations of the state intentionally harming its citizens were most prevalent.
Much criticism operates through low-risk, highly ambiguous signals. Thousands of messages consist of a government hashtag followed by only a single character. The numeral “6” (a homophone for liu 溜, meaning “slick”) and the character for niu 牛 (an abbreviation of niubi 牛逼, meaning “awesome”) function as ironic applause. Single periods (。) or ellipses (…) serve as textual equivalents to the blank sheets held in the subsequent street protests, representing messages that cannot be safely expressed. With a single character, users convey disbelief and exhaustion, signalling that they regard official explanations as transparent fabrications.
More elaborate forms of sarcasm expose the disconnect between official narratives and observable conditions to challenge the state’s authority to define reality. Rather than contesting state narratives directly, users satirize the authorities by attributing supernatural capacities to them or invoking fantasy biology. When the authorities announced just two days after the fire that there were no more new COVID cases, many joked that the prefectural authorities had exhibited “Chinese speed” (Zhongguo sudu 中国速度), or that the fire had scared away the virus.
Users frequently drew on shared historical and cultural knowledge to frame the one-party state as simultaneously predatory and in terminal decline, and therefore no longer worthy of public support. One asks: “Shouldn’t a numb, dark and blood-sucking government be overthrown?” The image of a parasitic government feeding on its citizens recurs throughout the corpus, frequently expressed through revolutionary writer Lu Xun’s 鲁迅 trope of a “cannibalistic society” (chi ren de shehui 吃人的社會).Footnote 60 Users also accuse officials of eating “steamed buns dipped in human blood” (renxue mantou 人血馒头) in the context of allegations of officials colluding with the Polymerase Chain Reaction (PCR) test industry to generate profits from mass COVID-19 testing. This image is taken from Lu Xun’s short story, “Medicine” (Yao 药), in which a father buys a mantou soaked in the blood of an executed revolutionary, believing it will heal his ailing son.Footnote 61
Many posts call for collective resistance, invoking revolutionary symbols to frame resistance as a patriotic duty. References to China’s National Anthem (“Arise, ye who refuse to be slaves”) and the May Fourth Movement recur frequently, sometimes comparing the regime with the Late Qing or Northern Warlord era. Such imagery portrays the party-state as a regime in its final days and challenges its self-portrayal as a modernizing force. A subset explicitly mobilizes for street protest, calling on citizens to join the White Paper protests,Footnote 62 to “occupy the government,”Footnote 63 or to “stand up” and “resist” (kangyi 抗议).Footnote 64 Others celebrate acts of resistance: onlookers confronting police,Footnote 65 students “speaking out,”Footnote 66 White Paper protesters “opposing tyranny.”Footnote 67 Demands for “freedom,” “democracy,” “human rights” and constitutionalism accompany accusations that the party-state has broken the social contract of political compliance in exchange for physical safety and basic livelihood. If the state has failed its obligations, users imply, citizens are released from theirs. Content of this kind is exceedingly rare on Chinese social media, not only because of censorship but also because inciting collective action carries the risk of severe punishment. The presence of such posts signals an escalation from policy critique to direct challenges to the party-state.
What these diverse strategies have in common is that users challenge a state they see as having departed from its ideals. By turning official rhetoric, revolutionary symbols and cultural authorities against the regime, they contest its right to rule on its own terms.
Contesting information control
A third category, approximately one-sixth of critical content, centres on contesting information control. Unlike the categories above, which target the state’s legitimacy or express emotional disaffection, these posts contest the state’s authority to establish the facts about the tragedy itself.
The struggle over narrative control began almost immediately after state media reported the disaster. In response to official claims that no obstructions had blocked emergency exits,Footnote 68 an anonymous blogger published a rebuttal that went viral, titled “The exits were clear, but they did not run” (lu shi tong de, tamen bu pao 路是通的,他们不跑). Using group chat logs and open-source data, the author alleged that residents were, in fact, barred from leaving the building.Footnote 69 Users subsequently demanded that state media investigate causes rather than just amplify official narratives.
Others confronted censorship through direct criticism. When a woman was swiftly detained for “spreading rumours about the number of fatalities,” users questioned the government’s priorities, implicitly contrasting its efficiency in silencing citizens with its failure to save them: “If only you were this quick in extinguishing fires.”Footnote 70 Some went further, turning the accusation of rumour-spreading back on the state: “Merely commenting is now perceived as ‘spreading rumours’ … You haphazardly assign nucleic acid test codes and skew data – aren’t these actions criminal?”Footnote 71
A more subtle tactic rendered censorship visible without directly naming it – for example, by highlighting the disconnect between high comment numbers displayed below messages and a scarcity of accessible content. One widely shared quote-post asked an apparently innocent question: “The comments [icon] indicate tens of thousands, but I can’t retrieve them. Is my Weibo malfunctioning?”Footnote 72 Others mocked the government’s demand for unquestioning loyalty through exaggerated sycophancy, displaying a compliance so excessive it became mockery: “Trust the government, indeed, indeed, indeed, good, good, good.”Footnote 73
As these examples show, users do not simply criticize censorship; they render it visible by contesting authoritative narratives, asserting their right to establish facts and mock-performing compliance to expose the state’s demand for uncritical belief. By disputing the state’s account of events and exposing its mechanisms of information control, users challenge the state’s monopoly on determining the truth.
Criticizing policy
The final category comprises direct critiques of China’s zero-COVID governance. Three distinct narratives recur: futility, disproportionate harm and excessive enforcement. Unlike systemic critique, which challenges the party-state’s right to rule, critique of policy failures targets the rationality and proportionality of specific measures. Some merely criticize sloppy policy implementation; others accuse the state of attempted murder.
Many highlight futility, not only because the SARS-CoV-2 Omicron variant was hyper-contagious but also because of the carelessness of state agents. One user remarked: “I’ve been isolating for 100 days. I’m safe. The only chance I have of becoming positive is when you people come to do a PCR test.”Footnote 74
Others argue that the human cost of containment exceeds its benefits, accusing the government of a continued implementation of a demonstrably harmful policy. Users conveyed their isolation by contrasting images of densely packed World Cup arenas with months of solitary confinement in small apartments and writing about children who had never known a world without masks, workers who had lost livelihoods and patients who had been turned away from hospitals. Users repeatedly claimed that the prevention regime had become more lethal than the virus: “The iron sheets couldn’t stop the virus, but they stopped ten lives.”Footnote 75 Although a handful of supportive posts warn against conflating pandemic measures with lethal accidents, the narrative that containment killed more than it saved prevails: “The girl who hanged herself, the mother who jumped from a building, the child who died because they weren’t rescued in time while sick.”Footnote 76
While critiques of futility and disproportionate harm are directed at policy outcomes, a third narrative ascribes suffering to a regime of domination operating through excessive policy enforcement.Footnote 77 As Xi Jinping had declared pandemic prevention to be a priority, cadres, bureaucrats and volunteers struggled to prevent new cases at all costs, resulting in measures that many felt undermined the policy’s stated purpose. Consequently, the regime created a principal-agent problem, where local enforcers pursued compliance metrics at the expense of protecting the public. Users documented the consequences of these incentives by highlighting physical obstructions, such as welded doors and iron wire seals, as well as the hard lockdown of neighbourhoods that blocked fire trucks and prevented residents from escaping the flames. Some described the testing regime as a form of ritualized subjugation: “We are forced [to go] down for PCR tests at 2 am in a blizzard. The children can’t sleep, and the elderly are freezing. Is this protecting us, or killing us?”Footnote 78
Taken together, these three narratives oscillate between criticizing specific policy failures and challenging the guiding logic of China’s zero-COVID regime. By framing the state’s policies as futile, harmful and coercive, users challenge the official pandemic control narrative. Instead of being a guardian of public health, the state is portrayed as the primary source of risk and suffering. Similar to the appropriation strategies discussed above, this inversion exploits the state’s own framing against it. While most of the criticism in this category is directed against pandemic policy, the boundary with systemic critique is fluid.
Discussion
Online mass events in China have become rare specimens for scholars of digital contention. A combination of technical filtering, disciplinary disincentives and the colonization of online public spaces by state-sponsored accounts has made online dissent a steep uphill battle. Without influential accounts to coordinate participation, initiate discussion and provide thematic focus, large-scale dissent would seem nearly impossible.
Yet, against all odds, OMEs occasionally do emerge. This study investigates how social media users overcame these obstacles to produce an OME following the Urumqi residential building fire. It reveals that users effectively self-mobilized by converging on official hashtags and embedding critical commentary en masse. This made the hashtags go viral, and dissent spread with them. The very infrastructure designed to contain dissent thus became its conduit. This article terms the mechanism through which this was achieved as “subversive amplification.”
While, in retrospect, it might seem self-evident as to why the Urumqi residential building fire would trigger mass dissent, other, similar events did not provoke such a reaction. The study’s results indicate that it was less the tragedy itself than a series of official communication failures that caused moral outrage and catalysed mass dissent: the fire chief’s blaming of the victims, the declaration that Urumqi was “basically virus-free,” the accusations of rumour-spreading. Thus, official messaging provided both the spark that triggered the outrage and the focal point on which it converged.
While hashtags facilitate convergence, they do not necessarily provide direction. Users hijacking state-sponsored Weibo messages were not united by a common grievance, and without opinion leaders to channel individual frustrations towards broadly visible collective demands, none emerged. In the brief time window available to them, users latched their criticisms onto the monolithic government narratives that dominate Chinese social media. This explains the heterogeneity of the content: not a single demand or overarching theme, but a catalogue of different emotions, frustrations and grievances. Official messages served as a Lennon wall onto which users pinned their grievances.
Whether users converged on a small number of hashtags strategically or reactively remains an open question. On one hand, it seems natural to react directly to the content that provoked them; on the other, users could have created their own hashtags or employed those of non-state accounts. Such behaviour can be characterized as semi-strategic: it was not coordinated but guided by a shared awareness of platform affordances.
The temporal proximity to the White Paper protests raises questions about online–offline linkages. There is little substantive evidence that online dissent on Weibo fuelled the subsequent protests. Very few messages mentioned either the protests or contained white paper symbolism. Given the risk of coordinating collective action in China, this absence is unsurprising. Nevertheless, the two events are characterized by virtually identical grievances and similar repertoires such as direct critique and creative symbolism. What role online spaces played in mobilization is a question future research could address.
There are several limitations. Results should not be extrapolated to Chinese society at large. Although participation was large enough to dominate Weibo, participants were a small, self-selected sample of the broader population. Moreover, as the literature review shows, creative strategies such as sarcasm, irony and literary allusion can serve as entertainment as much as evasion: a carnival in which users showcase creativity.Footnote 79 Furthermore, the findings, derived from a single case study, rely on a specific conjunction of conditions: a tragic event, state media engagement that created outrage and exploitable focal points, followed by a volume of self-mobilization sufficient to exploit platform affordances and overwhelm censorship capacity. Subversive amplification exploits a vulnerability in propaganda-saturated platforms, but it can do so only when the conditions are right. This explains why OMEs remain rare despite the mechanism being structurally available. Whether subversive amplification operates beyond such crisis moments, and whether the state has since adapted to address this vulnerability, are questions for future research. Finally, the study relies on data from a text-based social media platform; the results cannot be extrapolated to video platforms, social networks or private channels. Real-time collection of content avoided ex-post censorship bias but did not capture self-censorship and content removed before publication. Given that such content would likely have been more critical, the study’s findings are conservative.
The article extends existing research on digitally mediated contention in China by explaining how visibility can be achieved under tight information control. Its findings are also relevant for the broader scholarship on connective action and hashtag activism, which has theorized hashtags as focal points for loosely organized contention in democratic contexts. Subversive amplification represents an authoritarian variant: instead of creating grassroots hashtags, users appropriate those created by state-sponsored accounts. Despite its sophistication, the architecture of information control contains the seeds of its own subversion.
Supplementary material
Appendices are available online as supplementary material at https://doi.org/10.1017/S0305741026102562 and at https://phaidra.univie.ac.at/o:2344459.
Data availability statement
The data that support the findings of this study are available from the author upon reasonable request, subject to conditions that safeguard the privacy and safety of individuals in the dataset. The data are not publicly available as they contain personally identifiable information of Weibo users who posted on politically sensitive topics, and unrestricted disclosure could pose risks to their safety.
Acknowledgements
Participants of the research workshop “The Digital Reach of the Chinese State: Emerging Research Fields” in Hannover and the conference on “China: Ethics and Society” at Aarhus University provided valuable feedback to a previous version of this article. I am particularly grateful to Hanna Waight, King-wa Fu and the three anonymous reviewers for their helpful comments and suggestions. I would also like to thank the Bright Initiative by Bright Data for their technical support.
Funding information
The research for this article was funded by the European Research Council (Grant No. 678266).
Conflicts of Interest
None.
Christian GÖBEL is professor of contemporary China studies at the University of Vienna. His work focuses on the dynamic interplay between citizens testing the boundaries of participation and the state’s efforts to contain and channel these activities. He has led several major international research projects, including an investigation of local governance and citizen contention funded by the European Research Council.
Open access
