Normalization of Censorship: Evidence From China

Paper link

Introduction

• Recent scholarship claims that public awareness of censorship leads to backlash
• However, in China, censorship is widespread and overt
• Surveys show that Chinese citizens are apathetic towards or even supportive of censorship
• This applies even to people who've directly encountered censorship
• The literature focusing on backlash primarily focuses on censorship of government criticism and collective action
• However, authoritarian regimes censor far more than that
  • Hip hop music
  • Pop culture movies
  • Celebrity gossip
  • Pornography
• When censored content includes both politically threatening and seemingly harmless non-political content, citizens are less likely to view censorship as suppression of opposition
• Censorship becomes normal government policy
• This theory operates off the psychological effect of desensitization
• When both political and non-political content is censored, the act of censorship loses its ability to act as a signal that the government has something to hide
• Furthermore, this increases citizens' exposure to censorship, normalizing the act of witnessing censorship
• Data
  • Observational data from WeChat
    • 15872 censored articles
    • Only 40% are about collective action or government criticism
    • Majority are apolitical
  • Original survey experiment
    • Randomly expose respondents to varying degrees of non-political censorship
    • Respondents exposed to censorship of both political and non-political content are less outraged by censorship
• Challenges the existing theoretical frameworks that state that censorship in China is primarily targeted at political content and content that seeks to inspire collective action
• China censors everything
• This makes people numb to censorship
• Also challenges the conventional wisdom that overt censorship is more likely to lead to backlash than covert censorship
• Normalization of censorship implies that other coercive or authoritarian policies might become normalized
• Normalization can become another channel by which authoritarian governments maintain control

Normalization of Censorship: A Theory

Diluting The Proportion of Politically Threatening Content

• Backlash to censorship happens when citizens are aware of censorship and care about censored content
• Conventional wisdom holds that governments mostly censor politically threatening information
• Censorship is seen as an indication of government wrongdoing
• Streisand effect — trying to censor information draws more attention to it
• For this backlash to occur, citizens must believe that censorship is abnormal
• According to desensitization theory, when people's categorization of a stimulus shifts from negative to neutral, they no longer react as strongly to the stimulus
  • His citation on desensitization is a study on how exposure to violent movies and video games causes desensitization to violence in real life
  • Has that literature replicated?
• How do authoritarian regimes expose citizens to censorship in a normalized or even positive context?
  • Direct interaction with censorship (i.e. having one's own posts censored) is a negative experience
  • Indirect interaction, such as observing that a web page has been blocked, is more likely to be neutral
  • Incomplete censorship, such as allowing viewers to view the title of a censored story and its comment replies, allows users to infer the content of the censored topic and update their views accordingly
  • Allowing users to talk about censorship in a non-political context also dilutes the negative perception of censorship
    • Fans of Xiao Zhan, a pop star, often engage in online arguments that devolve into calls for the other side to be censored
    • Seeing censorship applied to non-political posts makes censorship seem less associated with repression
  • Only 9% of survey respondents said they'd been censored personally
  • 69.5% were aware of government censorship
  • Most citizens form beliefs about censorship via indirect exposure
  • By including non-political content in the range of censored topics, the government makes censorship seem neutral to those indirectly affected
  • If the topic being censored is sufficiently unpopular (such as pop culture among conservatives) censorship may even be viewed as a positive good

Increasing Citizens' Exposure to Censorship

• Censoring non-political content increases citizens' exposure to censorship
• Although initial exposure to censorship may provoke backlash, this reaction fades as people become more exposed to censorship
• This normalization is evident in how people call for censorship even in less serious, non-political contexts

Empirical Expectations

• Citizens exposed to normalized censorship will display less backlash against the act of censorship and the regime doing the censoring
• {$P(\textrm{non-political} | \textrm{censored})$} will be high

The Development of Censorship in China

• The range and scale of censorship in China is by far the largest in the world
• Earlier studies found that censorship in China largely targeted content likely to incite collective action while allowing most other forms of expression
• However, in recent years, overt censorship has become more aggressive
  • Ban seemingly harmless posts that include sensationalism, speculation or gossip
  • Censor content relating to the economy and business, which had previously been relatively uncensored
  • Extend censorship to apolitical fora, such as dating apps
  • More explicit intent to regulate online expression
    • Establishment of the Central Cyberspace Affairs Commission (CCAC) and the Cyberspace Administration of China (CAC) to centralize internet administration, including censorship
    • CAC regularly publishes reports on censorship activities
    • Justification for censorship is apolitical — government highlights negatives of an unregulated internet
    • Censorship is referred to with management terms rather than as a showcase of government power
• Censorship has also become more centralized, with increasing concentration of power in the CCAC

The Nature of Censored Content: Text Analysis

• Collect censored articles on WeChat from March 2018 to May 2020
• Classify censored articles into categories
• Determine the proportion of censored articles in each topic category
• 2 human classifiers, multiple text analysis models and cross-validation

Data Source

• WeChatScope — website created by a research team at the University of Hong Kong
  • Monitors 4000 WeChat public accounts
  • Accounts have a large number of subscribers and are therefore prime targets for censorship
  • Dataset has 15,872 censored articles from March 2018 to May 2020
  • I hope both the dataset and the researchers made it out
• WeChat is the single most popular social network in China
• Thus, when WeChat censors an article on a popular account, a large number of Chinese internet users will be exposed to the act of censorship
• WeChatScope is designed to record government censorship as traditionally defined (i.e. on political topics)
• This means that the expected bias in the dataset runs against the conclusion
• Dataset constitutes a "hard case" for the hypothesis
• However, one limitation of the WeChatScope data is that it only includes articles that have been subject to post-hoc censorship
• There are multiple layers of censorship, such as keyword filters and ensuring that the article is accessible through the Great Firewall
• Non-political content may have to pass a lower bar before being published
• However, Chinese internet users are adept at evading keyword filters
• Post-hoc censorship is the most extensive form of censorship on WeChat
• In addition, post-hoc censorship is the only form of censorship that will generate the normalization response, according to the hypothesis
  • What normalizes censorship is not seeing your own posts get censored, but seeing other people's posts get censored
  • This only occurs with post-hoc censorship

Categorization of Censored Articles

• Nine mutually exclusive categories
  • Political
    • Collective action
    • Government criticism
    • Government non-criticism
  • Non-political
    • Business and economy
    • Entertainment and sexuality
    • Advertisement
    • Local events, traditions and cultures
    • Foreign events
    • Other
• Categorization rubrics for non-political content explicitly excluded any content related to the government to avoid under-counting political content
• Furthermore, categories were applied in sequential order, with political categories applied first
• Ensures that ambiguously political articles are treated as political
• Methodology
  • Hand label 2500 articles
  • Use supervised text analysis to predict categories for remaining articles
  • Training set independently classified by two humans
  • Cohen's {$\kappa$} between human coders is 0.80, higher than the commonly used standard of 0.70
  • To predict the classification of the remaining unlabeled data, a multinomial logistic regression model with ridge estimator was trained on the 2500 labeled articles and run over the remainder of the dataset

Results

• Collective action and government related articles only accounted for 39.34% of all censored articles
• 60.67% of censored articles are apolitical
• The proportion of apolitical articles that were censored remained relatively constant over time
• Even during the COVID pandemic, when the government was going out of its way to censor criticism, the proportion of apolitical articles did not drop below 48%
• No clear pattern in censorship of non-political articles
• Majority of censored non-political content was from business & economy and entertainment & sexuality categories
• Expanding the definition of "political" to include vaguely related topics like the economy or foreign events only increased the proportion of censored political content to 57.15% — still a large proportion of non-political content that is censored
• The fact that collective action and government criticism form only a minority of censored content strengthens the government's case that censorship is benign
• No clear pattern for censorship of non-political content implies that censorship is broad and comprehensive
• This gives the government leeway to censor politically threatening content without drawing undue attention to it
• Also implies that Chinese citizens routinely encounter censorship of non-political content online, which produces favorable conditions for desensitization

Normalizing Effect of Censorship: Survey Experiment

• Experimentally manipulate censored topics and frequency of censorship
• Use ex-post evaluations of censorship policy and opinions regarding government to measure backlash

Participants

• N = 612
• Recruited using an online survey platform in China in December of 2020
• Political opinion surveys are relatively rare in China, and thus participants are not likely to be professional political survey takers
  • Sure, but are they professional survey takers?
  • I feel like many of the limitations of e.g. MTurk surveys also apply here
• Survey respondents were demographically similar to the broader population of Chinese internet users in terms of gender, urban/rural and region, but were richer and better educated
• However, interaction analysis finds no significant effect due to income or education

Experimental Design

• Compare reactions to "traditional" censorship, which blocks only government criticism and calls for collective action with "normalized" censorship that targets both political content and seemingly harmless apolitical content
• Three components
  • Pre-treatment questions collecting
    • Demographic data
    • Political interests
    • Economic ideology
    • Exposure to social media
  • Random assignment to control group or treatment group
    • Control group sees a sampling of stories with only political stories censored
    • Treatment group sees both political and non-political stories censored
  • Exposure to censorship
    • Both groups were asked to read 10 snippets of censored stories from the WeChatScope article database
    • Snippets were screenshots of real articles formatted in the same style as WeChat article previews
    • Six of the snippets were about non-political topics and four were about political topics
    • For the control group, three of the four political snippets had warning labels indicating that they were censored
    • The treatment group saw three censored political snippets and also three censored non-political snippets, testing both dilution and normalization effects
    • Snippets were validated by cross-checking them with a panel of China scholars, who agreed that none of them were absurd, fraudulent or otherwise unreasonable
• One big problem with this experiment design is that it relies on priming
• The way this experiment works is that study participants' opinions of censorship is primed by seeing only political stories censored, or both political and non-political stories censored
• This priming is then supposed to affect their opinion of censorship and of the government more broadly
• The problem is that many similar studies have failed to replicate

Measurement

• To measure backlash against the censorship policy, the survey participants were asked about
  • Their support for government control over the internet
  • Whether government control over the internet was normal
  • Whether they thought that the government was censoring too many articles
• To measure backlash against the regime participants were asked
  • About their assessment of the government
  • Overall satisfaction with China
  • Willingness in participate in protests
• Participants were asked about their satisfaction with the central government and local government separately, in an effort to defeat social desirability effects
• A Bonferroni correction was applied to p-values in order to compensate for the fact that multiple questions were being asked for the same hypothesis
• 9 covariates were used to check the balance between the treatment and control groups
  • Demographic covariates
    • Age
    • Income
    • Gender
    • Education
  • Political predispositions and internet usage covariates
    • Pro-market attitudes
    • Party membership
    • Political interest
    • Social media usage
    • VPN usage
• All covariates were balanced between treatment and control groups

Results

• Treatment increases respondents support for government and their belief that internet censorship is normal
• Exposure to censorship of non-political content reduces backlash against censorship of political content
• Exposing participants to non-political censorship increased support for both local and central government
• Large effect sizes given high baseline levels of support
• The fact that respondents are both more supportive of government and less willing to participate in protests indicates that the decline in willingness to participate in protests is a result of greater support for the government rather than intimidation from censorship
• Implications
  • Expanding the range of censorship reduces backlash
  • Repressive policies can increase regime support
  • Increasing awareness of censorship is not enough; people must be both aware of censorship and be educated about its repressive effect in order to oppose censorship

Limitations

• Limits to normalization theory
  • Does censoring any non-political content lead to normalization?
  • Is there a level of censorship that will cause a backlash even if applied to non-political topics?
  • Are there strategic considerations in choosing which non-political topics to censor?
• Limits to this study
  • Censorship exposure in the study was more intense than censorship as experienced by day-to-day WeChat users
    • Greater proportion of censored articles
    • Censorship was highlighted with a red warning label, whereas on WeChat users have to click on a headline in order to see that it has been censored
• How applicable are these results outside of China?
  • China has been under a censorious regime for a long time
  • Does this affect attitudes towards greater censorship?

Conclusion

• Existing literature claims that censorship awareness leads to backlash
• This fails to square with the reality of Chinese awareness of and apathy towards censorship
• This is because existing literature assumes that governments censor narrowly — only censor posts criticizing the government and advocating for collective action
• When the range of censorship is expanded to cover non-political topics, desensitization occurs
• This serves to explain why many examples of censorship generate more of an outcry in the West than they do in the authoritarian regimes in which the censorship takes place

• HomePage
• WikiSandbox

PmWiki

• Initial Setup Tasks
• Basic Editing
• Documentation Index
• PmWiki FAQ
• PmWikiPhilosophy
• Release Notes
• ChangeLog

pmwiki.org

• Cookbook (addons)
• Skins (themes)
• PITS (issue tracking)
• Mailing Lists

edit SideBar