How the Chinese Government Fabricates Social
Media Posts for Strategic Distraction, not Engaged
Argument∗†
Gary King‡

Jennifer Pan§

Margaret E. Roberts¶

May 23, 2016
Abstract
The Chinese government has long been suspected of hiring as many as 2,000,000
people to surreptitiously insert huge numbers of pseudonymous and other deceptive
writings into the stream of real social media posts, as if they were the genuine opinions of ordinary people. Almost all scholars, activists, journalists, and participants
in social media claim these so-called “Fifty Cent Party” posts argue vociferously for
the government’s side in political and policy debates. Yet, almost no systematic empirical evidence exists for this claim, or, more importantly, for the Chinese regime’s
strategic objective in pursuing this activity. In the first large scale empirical analysis of this operation, we show how to identify the secretive authors of these posts,
the posts written by them, and their content. We estimate that the government fabricates and posts about 488 million social media comments a year. In contrast to prior
claims, we show that the Chinese regime’s strategy is to avoid arguing with skeptics
of the party and the government, and to not even discuss controversial issues. We
infer that the goal of this massive secretive operation is instead to regularly distract
the public and change the subject, as most of the these posts involve cheerleading
for China, the revolutionary history of the Communist Party, or other symbols of
the regime. We discuss how these results fit with what is known about the Chinese
censorship program, and suggest how they may change our broader theoretical understanding of “common knowledge” and information control in authoritarian regimes.

∗

Our thanks to Neel Guha, Peter Dyrud, Yingjie Fan, and many others for superb research assistance;
Peter Bol, Becky Fair, Chase Harrison, Blake Miller, Jean Oi, Samantha Ravich, Andy Walder, Yuhua Wang,
Chaodan Zheng, and Yun Zhu for helpful comments; and DARPA (contract W31P4Q-13-C-0055/983-3)
and the National Science Foundation (grant 1500086) for research support.
†
Authors are listed alphabetically.
‡
Albert J. Weatherhead III University Professor, Institute for Quantitative Social Science, 1737 Cambridge Street, Harvard University, Cambridge MA 02138; GaryKing.org, King@Harvard.edu, (617) 5007570.
§
Assistant Professor, Department of Communication, 450 Serra Mall, Building 120, Stanford University,
Stanford CA 94304; http://people.fas.harvard.edu/∼jjpan/, (917) 740-5726.
¶
Assistant Professor, Department of Political Science, University of California, San Diego, Social Sciences Building 301, 9500 Gilman Drive, #0521, La Jolla, CA 92093-0521, meroberts@ucsd.edu, MargaretRoberts.net.

 1

Introduction

Social media in China appears as vibrant and extensive as in any Western country, with
more than 1,300 social media companies and websites, and millions of posts authored
every day by people all over the country. At the same time, the Chinese regime imposes
extensive controls over of the entire system. Which social media companies are prevented
from operating in China is easy to see (the so-called “Great Firewall of China”), and the
scholarly literature now offers considerable evidence on how and why they censor certain
individual social media posts that have appeared on the web or filter them out before
appearing. In both cases, the censorship apparatus allows criticism of the regime, its
officials, and their policies (which can be useful information for the central government
in managing local leaders) but stops discussions that can generate collective action on the
ground (King, Pan and Roberts, 2013, 2014).
According to numerous speculations by scholars, activists, journalists, officials in
other governments, and participants in social media, the Chinese regime also conducts
“astroturfing,” or what we might call “reverse censorship,” surreptitiously posting large
numbers of fabricated social media comments, as if they were the genuine opinions of
ordinary Chinese people. The people hired for this purpose are known formally as “Internet commentators” (网络评 论员) but more widely as “50c party” members (五毛党),
so-called because they are rumored to be paid 50 cents (5 jiao, 角, or about US$0.08) to
write and post each comment. (Of course, political parties do not exist in China and so,
despite the name, the “50c party” is not a political party.)
The nearly unanimous view of those who have written about this subject is that 50c
party propaganda argues with and debates against those who criticize the government, its
leaders, and their policies. (This was our view as well, prior to the research reported here.)
In Section 2, we systematically summarize these views, including a quantitative analysis
of a body of social media posts accused of being written by 50c party members. We show
that all these sources agree on the purpose of the 50c party. However, until now no method
has existed for detecting 50c party members, which posts they write, their content, or why
they write them, and so the best anyone could do was to base these claims on intuition,

1

 logic, a small number of available anecdotes, rumors, and leaked government directives.
In this paper, we offer the first systematic empirical evidence for the content of 50c
party posts and the government’s strategic objectives.1 We begin by analyzing an archive
of emails leaked from the Internet Propaganda Office of Zhanggong, a district of Ganzhou
City in Jiangxi province (Section 3). These emails give explicit details of the work of
numerous 50c accounts in this district. Although in the public domain and reported on
in the press (e.g., Henochowicz, 2014; Sonnad, 2014), the structure of this archive is
complicated, too large to understand by traditional qualitative methods, and in formats
(and attachments) far too diverse to make standard methods of automation feasible; as
such, it has never before been systematically analyzed, and little of it has been explored.
We have developed an approach to analyze this data set and, from it, have extracted more
than 43,000 known 50c party posts and their authors.
We first characterize these emails, social media accounts, and known 50c posts via
their network and time series structure. Then, in Section 4, we systematically analyze
the content of the 50c posts in our leaked archive and, in stages, extrapolate to the rest
of China. To do this, we first obtain the Sina Weibo accounts of the 50c party members
revealed in the leaked emails and obtain from the web many other social media posts they
authored. Next, based on measurable features of the authors, their accounts, and the posts
in our verified Zhanggong sample (but not their content), we show how to identify 50c
party members and their posts throughout China. We then use this methodology to study
the content of the posts and finally ascertain the goals behind this massive government
program and how it fits with and may reveal broader government strategies.
We validate our 50c party member predictions in Section 5 with a novel sample survey
of predicted 50c party members, with several gold standard evaluations (unusual for any
survey), to validate the survey and our predictions. In Section 6, we estimate and reveal the
size of what turns out to be a massive government operation which writes approximately
488 million posts a year.
At every stage, our results indicate that the prevailing view of the 50c party is largely
1

Because China is a single-party regime, we use the terms “government” and “regime” interchangeably
to refer to those in power.

2

 incorrect. We show that almost none of the Chinese government’s 50c party posts engage
in debate and argument of any kind. They do not step up to defend the government, its
leaders, and their policies from criticism, no matter how vitriolic; indeed, they seem to
avoid controversial issues entirely. Instead, most of these posts are about cheerleading
and positive discussions of valence issues which, we infer, is a strategy designed is to
actively distract and redirect public attention from ongoing criticism, other grievances, or
collective action.
Section 7 gives a unified parsimonious summary of Chinese government internal information control, explores the logic behind it, and shows how these findings may cause
scholars to rethink the notion of “common knowledge” in theories of authoritarian politics
more generally. Section 8 concludes and then gives a summary of what we might have
missed and how scholars can follow up on this work.

2

What We Think We Know

We summarize here the common views of (1) academics, (2) the news media, and (3)
social media participants, each of whom characterize the 50c party as engaging in “handto-hand” verbal combat, making specific, directed arguments that support the government,
its leaders, and their policies, and opposing arguments to the contrary by engaging in
debate and criticism of its enemies, including those who oppose it within China and from
abroad. Under (3), we also introduce and validate a scheme to categorize 50c posts on
the web; we use it in this section to understand the posts accused of being written by 50c
members and for many other purposes throughout this paper.
Although the work cited in this section involves considerable effort and creativity on
the part of the authors, the difficulties of collecting data on an inherently secret operation
means that prior literature includes “no successful attempts to quantify regime-sponsored
commentary in China” (Miller, 2016), and even a few clever efforts to guess the values of
dependent variables before turning to try to explain or predict the guesses (Miller, 2016;
Han, 2015). Indeed, even sophisticated unsupervised statistical techniques have produced
“no evidence of large-scale Wumao [50c] activity on Weibo” (Yang, Yang and Wilson,

3

 2015). As most authors acknowledge, claims about the content and extent of 50c party
posts rest on little or no empirical basis.
1. Academics Academics report that between 250,000 and 300,000 paid 50c party
members write posts as directed by the Chinese government (Freedom House, 2009; Barr,
2012; Greitens, 2013), so called because “rumor has it that they receive 50 cents for each
post supporting the government” (Tong and Lei, 2013). The 50c party members “mix
control and activism on line. . . making favorable comments, and generally pushing discussion toward pro-Party lines” (Greitens, 2013). They “patrol chatrooms and online forums, posting information favorable to the regime and chastising its critics” (Deibert and
Rohozinski, 2010). They are an “army of online commentators. . . promoting the Chinese
Communist Party’s line on sensitive subjects” (Bremmer 2010, see also Hassid 2012).
They “faciliate state propaganda and defuse crises” (Han, 2015).
2. Journalists

Journalists and others express very similar views, describing 50c mem-

bers as “undercover pro-government Internet commenters” (Keating, 2011) who “set out
to neutralize undesirable public opinion by pushing pro-Party views through chat rooms
and Web forums” (Bandurski, 2008). They “shape online public opinion” by labeling
“critical opinion leaders as traitors of the country” (Lam, 2012). Prominent dissident Ai
Weiwei said “If you oppose the US and Japan [online], you are a member of the 50 cents
army” (Strafella and Berg, 2015). 50c party members “combat hostile energy,” defined
as posts that “go against socialist core values,” or “are not amenable to the unity of the
people.” Such information should be “resolutely resisted, proactively refuted, and eagerly
reported to Internet authorities.” (Haley, 2010). Through active engagement of opposition views, they try to “sway public opinion” (Editors, 2016; Ng, 2011), “influence public
opinion. . . pretended to be ordinary citizens and defending or promoting the government’s
point of view” (Lam, 2013), or “steer conversations in the right direction” (Editors, 2013).
Perhaps the only difference between the journalists and academics is that some journalists
report much higher numbers of 50c party members, ranging from 500,000 to 2 million
(Philipp, 2015).

4

 3. Social Media Participants In addition to academics and professional journalists,
participants in social media also characterize 50c party members by openly accusing some
of being members themselves. To understand their views, we obtained a random sample
of 9,911 social media posts from 2010 to 2015 that contain the word “五 毛 党” (“50c
party”). From these data, we drew a sample of 128 posts written by people accused in
other posts of being 50c party members.2
We then sorted these “accused 50c posts” into one of six categories, using a categorization scheme we will use throughout this paper. With two independent Chinese language
coders, and 200 randomly selected posts, we measured the inter-coder reliability of the
categorization scheme at 93% agreement (see the Appendix A for details). Two of the
categories, comprising 65% of the accused 50c posts, represent the views of academics
and journalists, and include (1) Taunting of Foreign countries (which is 29% of this sample) and (2) Argumentative praise or criticism (36% of the sample). Taunting includes
favorable comparisons of China compared to other, usually Western, countries and taunting of pro-democracy or pro-West values or opinions. Argumentative praise or criticism
involves engaged argument and debate about controversial (non-valence) issues, criticism
of opponents of the government, or praise of the leaders.
The categorization scheme also includes (3) Non-argumentative praise or suggestions
(22% of the sample) and two categories that everyone agrees are not what 50c party
members are writing about, (4) factual Reporting (8%) and (5) Cheerleading (at 5%).
Non-argumentative praise or suggestions includes discussion of noncontroversial valence
issues, such as improving housing or public welfare, or praise of government officials, but
does not debate or take opposing viewpoints. (Category (3) does not threaten the regime
in any way, and indeed Chen, Pan and Xu (2016) show that local governments openly
discuss non-argumentative valence issues with others on government web sites.)
Factual reporting involves descriptions of government programs, events, initiatives, or
2
Accused 50c party members are distinct from “volunteer 50c members” (自干五; literally “self do
50c” also translated as “bring your own grain”), who express pro-regime or anti-western sentiment online
without being paid by the government, the “little red flowers” (小粉红), an unpaid red guard who also attack
opponents of the regime online, and the “American 50c” (美分党) who express western democratic values
and criticize the Chinese communist regime online. None of these are known to be organized groups.

5

 plans. Cheerleading includes expressions of patriotism, encouragement and motivation,
inspirational slogans or quotes, gratefulness, discussions of aspirational figures, cultural
references, or celebrations. (Appendix A also includes a sixth “other” or “irrelevant”
category, but we remove this so that the percentages from the first five categories add to
100%.)
Thus, social media participants accusing others of being 50c party members agree
with scholars and academics that the content of 50c posts is basically antidisestablishmentarianism — arguing with those who oppose with the regime, its leaders, or their
policies.
We now go a final step and study the identities of accused 50c party members, which
can be difficult because such accusations occur on comment or discussion threads where
participants are anonymous. However, by carefully cross-referencing profile information
across multiple platforms, we were able to unearth personal details for a handful of individuals accused of being 50c members. The background of these individuals vary greatly,
but in each case, it is unlikely that they are being paid to make posts at the behest of
the government. For example, among those accused of being 50c party members include
Zhou Xiaoping (周小平), a blogger well known for his anti-west and nationalist sentiment, and He Jiawei (何家), a blogger known for critiques of the Chinese government
who posts on Boxun, a site hosted outside of China devoted to covering topics such as
Chinese government human rights abuses. Other well known figures accused of being
50c include Lin Yifu (林毅夫), a Peking University professor who was Chief Economist
and Senior Vice President of the World Bank from 2008 to 2012. However, those accused
of being 50c party members also include figures not connected to politics such as, in our
data, a comedian, a lawyer, and a marketing executive.
It appears that the evidence base of those accusing others of being 50c party members
is no better than that of academics or journalists. Although the prior beliefs of all three
groups about the content of 50c party posts are almost the same, little evidence supports
their claims.

6

 3

Leaked Internet Propaganda Office Communications

The problem in the literature has been that “detecting the Wumao [50c party] is difficult
because there is no ground truth information about them” (Yang, Yang and Wilson, 2015).
We are fortunate to be able to change this situation. In December 2014, anonymous
blogger “Xiaolan” (https://xiaolan.me/) released an archive of all 2013 and 2014
emails to, and some from, the account of the Internet Propaganda Office (网 宣 办), a
branch of the propaganda department, of Zhanggong district. Zhanggong district is a
country-level administrative unit (with a population in 2013 of 468,461) that is part of
the moderate sized Ganzhou City, in Jiangxi province. The emails reported activities of
Internet commentators, including numerous 50c posts from workers claiming credit for
completing their assignments, and many other communications. The hack was widely
reported and the archive of emails have been publicly available since (e.g., Henochowicz,
2014).
The archive’s large size, complicated structure, numerous attachments, diverse document formats (screen shots, Word, Excel, Powerpoint, raw text, text as part of other
emails, etc.), multiple email storage formats, and many links to outside information has
made digesting much of it impossible either for individuals reading and coding by hand
or for existing methods of automated text analysis. Journalists managed to pull out a few
examples to write newspaper articles, but no systematic analysis has been conducted of
these data.
To systematize this richly informative (and essentially qualitative) data source, we developed and applied a variety of methods and procedures, from large scale hand coding,
to specially tuned and adapted methods of named entity recognition, to methods of automated text analysis and extraction. Because of the considerable effort and resources
necessary, we will make structured and easy-to-access forms of these data publicly available so that others may follow up. Full replication information for all analyses in this
paper will be available as well.
From this work, we identified 2,341 emails sent from 2/11/2013 to 11/28/2014. Of
these, 1,245 contained the text of one or (usually many) more 50c posts. In all, we har-

7

 vested from these emails and their attachments 43,797 known 50c posts that form a basis
for our subsequent analyses and, as a training set, help identify other 50c posts. (Although
we have the name, direct contact information, and often photographs of many of the people discussed in this paper, we have no academic reason to make this information more
public than it already is; we therefore do not do so.) We conduct rigorous evaluations of
our claims to the next section. For now, we characterize the content with three separate
descriptive analyses.
First, we portray the overall structure of communications in these emails with the
network diagram in Figure 1. Each circle is a specific email account and each line denotes
where one or more emails was sent from and to. The large flower-like shape at the bottom
are 50c party members sending in copies of their posts to the Zhanggong district Internet
Propaganda Office (章贡区网宣办), claiming credit for completing their assignments.
This office then reports up to other offices (see the lines out from the center of the flower
shape), These include the speaker of Zhanggong People’s Court News office (江西省赣
州市章贡区人民法院新闻发言人), and District Party Office Information Department
(区委办信息科).
Second, although the scholarly literature describes 50c party members as ordinary
citizens hired for very low wages, we found instead that almost all in our sample are government employees.3 Of the 43,797 posts, only 291 were made by individuals or groups
we could not identify (the content of these posts were very similar to those we could identify). The remaining 99.3% were contributed by one of over 200 government agencies
throughout the Chinese regime’s matrix organizational structure (of geographic representation by functional area) in Zhanggong district. This included 9,159 posts (20.1% of the
43,797 total) made directly by the Zhanggong Internet Propaganda Office, 2,343 (5.3%)
by the Zhanggong district Bureau of Commerce (区 商务局), 1,672 by Shuixi Township
(水西镇, one of several townships in Zhanggong), 1,620 by Nanwai sub-district (南外街
道 (one of several sub-districts in Zhanggong). Others come from functional bureaus in
Zhanggong district (e.g., 体育局 Sports Bureau, 人保局 Bureau of Human Resources and
3

Han (2015), who seems to have gotten it right, argued that internet commentators are often public
servants though some may be specially recruited.

8

 Figure 1: Network Structure of Leaked Email Correspondents. Circles are email correspondents, edges (lines) indicate email correspondence. Most of the correspondence is
in toward the center of the flower-like structure (to the Zhanggong Internet Propaganda
Office and then out from that office to higher level offices.
Social Security, 地税局 Bureau of Taxation, 法院 Zhanggong district court), the government offices of Zhanggong’s sub-districts and townships (e.g., 沙河镇 Shahe Township,
赣江街道 Ganjiang sub-district), functional departments in each sub-district or township
(水西镇党政办 Shuixi Township Party Office), and administrative offices of neighborhoods and villages in Zhanggong’s townships and sub-districts (e.g., 南外街道东阳山
社区 Dongyang Shan neighborhood of the Nanwai sub-district, 水西镇何 乐村 Hele
village of the Shuixi sub-district).
Of the 50c posts in this archive, 29.98% did not contain a URL or a description of the
site where the content was posted. Of the remainder, 53.54% of the 50c posts were com-

9

 ments on government sites (GanzhouWeb, Newskj, DajiangWeb, JidanWeb, JiangxiWeb,
CCTVWeb, RenminWeb, JiujiangWeb, QiangGouWeb), and 46.67% were on commercial sites. Of the 50c posts on commercial sites, 54.06% were on Sina Weibo, 32.05% on
Tencent Weibo, 10.73% on Baidu Tieba, and 2.69% on Tencent QZone, with the rest in
the long tail receiving less than 1% each.
We also found no evidence that 50c party members were actually paid fifty cents or
any other piecemeal amount. Indeed, no evidence exists that the authors of 50c posts
are even paid extra for this work. We cannot be sure of current practices in the absence
of evidence but, given that they already hold government and Chinese Communist Party
(CCP) jobs, we would guess this activity is a requirement of their existing job or at least
rewarded in performance reviews.
Finally, we offer a first look at the content of the 43,797 posts from the 50c party that
we unearthed. We do this by plotting in Figure 2 a daily time series of counts of these
posts. The most important finding in this graph is that the posts are far from uniformly
distributed, but instead are highly focused into distinct volume bursts. This suggests a
high level of coordination on the part of the government. Indeed, often the most influential
patterns on social media are the bursts that occur naturally when discussions go viral. The
government’s manufactured bursts mirror these influential naturally occurring patterns.
Bursts are also much more likely to be effective at accomplishing specific goals than a
strategy of randomly scattering government posts in the ocean of real social media.
Although we will conduct rigorous, quantitative analyses of the content of 50c posts
in the sections to follow, we also provide a feel for the general content and time line of
these posts by labeling the largest volume bursts in this set. The labels are brief qualitative
summaries we chose from reading numerous posts.
As with most social media time series plots like this in most countries, regardless of
classification, the 50c posts in this collection cluster in volume bursts. We thus describe
each volume burst (with numerical labels corresponding to those in the figure). As can
be seen throughout, this is the first indication that the overwhelming focus of these posts
is on cheerleading and distraction, rather than engaged argumentation and debate. The

10

 4000

1. Qingming
festival
(April)

2000

6. Urumqi rail
explosion (May)

3. Shanshan
riots (July)
2. China
Dream
(May)

1000

Count of Posts

3000

8. Martyr's
Day
(Oct)

7. Gov't
forum,
praise
central
subsidy
(Jul−Aug)

5. Two meetings
(Feb)

0

4. 3rd plenum
CCP 18th
Congress (Nov)

Jan

Mar

May

Jul

Sep

Nov

Jan

Mar

May

Jul

Sep

Nov

Date (Jan 2013 − Dec 2014)

Figure 2: Time series of 43,797 known 50c social media posts, with qualitative summaries
of the content of volume bursts.
timing some of these bursts may even suggest a possible government strategy of following
collective action events with posts to distract the public with an interesting, but innocuous
and unrelated topic.
1. Qingming (Tomb Sweeping Day): over 18,000 posts about veterans, martyrs, how
glorious or heroic they are, and how they sacrificed for China.
2. China Dream: Over 1,800 posts about President Xi Jinping’s “China Dream”. Potentially a reaction to the April 2013 People’s Daily piece instructing municipal
governments to carry out “China Dream” propaganda campaigns (see http://j.
mp/chinadream).
3. Shanshan Riots: 1,100 posts, immediately following Shanshan riots in Xinjiang. At
5:30pm Zhanggong county sent an email to itself (probably BCCing many others),
highlighting three popular posts about Xinjiang, and saying this was a terrorist incident. At 8:00pm on the same day, Zhanggong sent an email to Ganzhou City to
11

 which it reports with hundreds of 50c posts about China Dream, local economic
development, etc.
4. 18th Party Congress, 3rd Plenum: Over 3,400 posts related to the 3rd plenary session of the Chinese Communist Party’s 18th Congress, which discussed plans for
deepening structural reform.
5. “Two Meetings”: Over 1,200 posts about Ganzhou’s People’s Congress and Political Consultative Committee meetings, and policies to be discussed at the two
meetings, including factual reporting of environmental issues, one child policy, rural issues, growth and development.
6. Early May Burst: 3,500 posts, about a variety of topics, e.g., mass line, two meetings, people’s livelihood, good governance. Immediately followed the Urumqi Railway Explosion, but no connection other than timing is apparent.
7. Over 2,600 posts celebrating the second anniversary of “Central Soviet Areas Development policy” (若干意见), subsidies from the central government to promote
the development of region where the original CCP bases were located (including
the region where Zhanggong is located); at the same time, the local government
held an online Q&A session for citizens.
8. Martyr’s Day: 3,500 posts about martyrs and the new Martyr’s Day holiday, celebrating heros of the state.
We now turn to a more systematic analysis of these posts, their accounts, and others
like them beyond Zhanggong.

4

Content of 50c Posts

We now reveal the content of our 50c party posts from across China by estimating the
distribution of these posts across the five main content categories introduced in Section 2
(with details in Appendix A). We do this in five separate analyses and data sets which
successively expand the target set of 50c party posts to larger and larger areas across
12

 the country, as much as possible making ourselves vulnerable to being proven wrong.
Through each of these analyses, we find very similar results, with 50c party posts largely
comprised of cheerleading and distraction rather than engaged argument.
1. Leaked 50c Posts

We first analyze the 43,797 50c social media posts we harvested

from numerous authors and different social media sites in the leaked archive from Zhanggong. To study these data, we first hand code a random sample of 200 posts into our six
categories, with high inter-coder reliability.
One result is immediately apparent: The number of posts from this sample that fall
in the categories “taunting of foreign countries” or “argumentative praise or criticism” is
exactly zero. This is an important surprise, as it is essentially the opposite of the nearly
unanimous views espoused by scholars, journalists, activists, and social media participants. This result would be highly unlikely to have resulted from (Binomial) sampling
error if the true share of the full set were even as large as a few percentage points (at 5%,
which would still be a major surprise, the probability of seeing the sample we obtained is
essentially zero). To push even farther, we did extensive searches and reading among the
remaining posts and finally found a few that fit this category (see the examples in Categories 2 and 3 in Appendix A), but the overall result is that 50c party posts are extremely
rare in these categories.
We thus infer that the leaked posts contain very little Taunting of Foreign Countries
or Argumentative Praise or Criticism, but we go ahead and verify by formally estimating all the category proportions in the entire set of posts. Using a text analytic method
known colloquially as “ReadMe” (named for the open source software that implements
it), we estimate the category proportions directly, without having to classify each post into
a category (Hopkins and King, 2010). This is fortunate, since individual classifiers that
manage to achieve high (but imperfect) levels of the percent correctly classified, may still
generate biased estimates of the category proportions. For example, an estimate indicating
that zero country dyad-years, since WWII, were at war achieves a predictive accuracy of
about 99.9%, but aggregating these classifications yields an obviously biased (and useless)
estimate of the prevalence of war. In contrast, ReadMe does not give individual classifi13

 cations, but it has been proven to give approximately unbiased and consistent estimates
of the category proportions, which here is the relevant quantity of interest. The other advantage of ReadMe in this context is that its statistical assumptions are met exactly by our
sampling procedures.
The estimated proportions of 50c posts by category for all data sets appear in Figure 3;
the results for our first data set (of all posts found in the leaked emails in Zhanggong) are
represented by a histogram, formed by the set of solid disks ( ) for the point estimate,
and solid line for the confidence interval, for each of the categories. Other results, to be
described below in order from left to right within each category, also appear in the same
graph.
The categories in Figure 3 are arranged so that the two emphasized in the literature
appear on the left, and our main empirical results on the right. For this analysis, the
results indicate that approximately 80% fall within the Cheerleading category, 13% in
Non-argumentative praise or suggestions, and only tiny amounts in the other categories,
including nearly zero in Argumentative praise or criticism and Taunting of foreign countries. Clearly, these results clearly indicate that 50c posts are about cheerleading, not
argumentation.
2. Posts from Leaked 50c Weibo Accounts One possibility we now consider is whether
50c party members differentially reported cheerleading posts back to the Propaganda Department, even though they posted about topics at the behest of the regime from other
categories as well. To study this question, we construct a second data set by first identifying all Weibo social media accounts revealed in the leaked email archive. We choose
Weibo because it is the most widely used social media site that enables mass distribution,
and we were able to obtain access in the manner we needed it. We then found these accounts on the web and kept all 498 Weibo accounts that made at least one post. Finally, we
downloaded all social media posts from these accounts, yielding a set of 167,971 known
— but not previously leaked — posts from 50c accounts.
We drew a random sample (stratified by account) of 500 of these 167,971 social media posts, and coded them into our categories as a training set. In this randomly selected
14

 1.0

Leaked e−mails, all sites
Leaked accounts, Weibo
Leaked accounts, ordinary
Leaked accounts, exclusive
Within county prediction, all posts
Out of county prediction

Cheerleading

0.6

●

0.4
0.0

0.2

Proportion

0.8

●

Argumentative Praise
or Criticism
Taunting of Foreign
Countries
●

●

Factual
Reporting

●
●

Non−argumentative
Praise or Suggestions

Figure 3: Content of Leaked and Inferred 50c Posts, by substantive category (with details
in Appendix A) and analysis (given in the legend)
training set, like the last, we find no evidence of taunting of foreign countries, although we
did find a handful of posts in the category of Argumentative Praise or Criticism, constituting only 3% of the posts. As above, we then used (a stratified sample and) ReadMe to
estimate the five category proportions for the set of all posts. The results, reported in the
second bar of the histogram in Figure 3, are very similar to that from the first data set. The
point estimates (portrayed as solid triangles, , with confidence intervals as dashed lines)
indicate that again the bulk of 50c posts from leaked accounts are Cheerleading (51%),
22% in Factual Reporting, 20% in the category Non-argumentative praise or suggestions,
and only 0.4% in the category Argumentative praise or criticism.
3. Partitioning Leaked Accounts for Extrapolation

We designed our third analysis

to further understand the leaked data and also to prepare the ground for extrapolation —
to previously unleaked accounts in Zhanggong in Analysis 4 and to other counties across
China in Analysis 5.
We first studied the structure of the 498 known 50c Weibo accounts and their 167,971
15

 social media posts. What we found are two fundamentally different types of accounts. In
both, we often find many commercial posts, which fall in our “other” category (see the
Appendix). Since we remove and condition on this “other” category for all analyses, we
do not define account types on this basis either. Thus, the first type of account, which
we call ordinary, are used by apparently ordinary people in China to post about their
children, funny videos, commercial advertisements, sports teams, pop stars, and many
other subjects. Embedded within the stream of these posts are others which these authors indicate in their communication with the Propaganda Department to be 50c party
posts. The second, which we call exclusive accounts, are (aside from commercial posts)
almost exclusively devoted to 50c posts, as designated by email communications with the
Propaganda Department in our leaked archive.
Distinguishing between ordinary and exclusive accounts in our leaked archive is easy
(the percent of real 50c posts reported to the Propaganda Department in each is a direct
measure), but our goal is to extrapolate to other counties where we have no known 50c
posts. Thus, we need a partitioning algorithm that will identify accounts without the inside
information in our extraordinary leaked data. Moreover, since our goal is to determine the
content of 50c posts, we must be able to discern whether an account is owned by a 50c
party member without using the text of the posts.
To develop these rules, we followed the logic of “Bayesian falling rule list” methodology, which are accurate and also highly interpretable (Letham et al., 2015). The interpretability also enabled us to combine qualitative knowledge with modern machine learning, as well as to make choices that were much easier to apply outside of Zhanggong.
With this approach as a guide, we found the following two rules to partition our 498 50c
accounts, and which apply more generally: First, we obtain candidate 50c accounts by collecting all accounts that comment or forward on any post on the Zhanggong government
weibo account (http://weibo.com/u/3880516376). Second, we narrow this to
accounts with 10 or fewer followers. These two simple, interpretable rules are highly
plausible and consistent with what is known about social media. After all, accounts that
engage with government websites and have no more than a handful of followers were

16

 surely created or at least used for a very narrow purpose.
Exclusive accounts will obviously be considerably easier to find outside of our leaked
archive than ordinary accounts, and so we would like to use this fact to narrow our search
criterion. However, if the two types of accounts post different content, then using this
simplifying rule will also induce bias. We thus evaluate this potential problem in the data
where we can verify it. Hence, we begin by using the two rules and a third that narrows
the results to only the 498 known 50c accounts in our archive. We then partition these
known 50c Weibo accounts into the two categories. The result is that we find that 202
(41%) are exclusive accounts and the remaining 296 (59%) are ordinary accounts.
This partition of the data is neither right nor wrong (and so statistics like “percent
correctly classified” do not apply), but it is useful only to the extent that using only the
exclusive posts causes no bias. Thus, we estimate and compare the distribution of posts
within the ordinary and exclusive account types across our five content categories. To do
this, we apply ReadMe within each partition and compare the results. Fortunately, the
results are very close to each other, and (as a result) to the overall results we presented
above. This means that no bias would be induced by narrowing our search outside our
leaked archive to exclusive accounts.
Point estimates for the category proportions appear in Figure 3 (marked as

and

,

in red). For both, the bulk of 50c posts appear in the Cheerleading category (46% for
exclusive accounts and 58% for ordinary accounts). In contrast, the sum of taunting of
foreign countries and of argumentative praise or criticism is very small (4% for exclusive
and 10% for ordinary).
4. Unleaked 50c Posts in Zhanggong We now use the results of Analysis 3 and expand
our study of 50c posts beyond the leaked archive. In this section, we focus on previously
unidentified 50c posts in Zhanggong. To do this, we choose exclusive accounts (by applying the two rules from the previous section). With this procedure, we find 1,031 accounts,
of which 829 accounts are not mentioned in our leaked archive. We then find and scrape
all 22,702 social media posts available from the front page of each of these accounts.
Each front page has up to 45 separate posts. We then analyze these posts with ReadMe,
17

 as above.
Results from this analysis appear in Figure 3 (with point estimates represented by ×).
The result again is very similar to previous analyses: 55% of the posts made on these
accounts engaged in Cheerleading, 17% engaged in Factual Reporting, 24% engaged in
Non-argumentative praise and suggestions, and about 4% in Taunting of foreign countries
and essentially zero in Argumentative praise or criticism.
5. Unleaked 50c Posts in Counties with County Government Weibo Accounts

We

now extrapolate to counties across China. We do this by starting with all 2,862 counties
(and county-level divisions). We then take as our target of inference 50c behavior in
1,338 of these counties which are structured same way as Zhanggong, with a Propaganda
Department that has a public web site. We then drew a simple random sample of 100
of these counties and identified all exclusive accounts and a sample of their social media
posts.4
To be more specific, for each county government Weibo account, we collected all
151,110 posts, randomly sampled up to 200 posts of these, identify all outside Weibo
accounts that commented on or forwarded any one, downloaded all meta-data from those
accounts, and subsetted to those with 10 or fewer followers. We then download the first
page, comprising up to 45 social media posts from each account, as our candidate 50c
posts.
Figure 3 provides our results (with point estimates represented as a diamond, ).
Again, we find very similar results, highly focused on cheerleading and distraction rather
than argumentation and criticism: 64% of the posts made on these accounts are categorized as Cheerleading, 21% in Factual Reporting, 8% Non-argumentative praise and
suggestions, 0.2% in Taunting of foreign countries, and only 1% in Argumentative praise
or criticism.
4

Many of the remaining 1,524 counties have Weibo accounts run by government bureaus and agencies
(such as public security department, civil affairs department), but not by the county-government. Our
informal study of these counties revealed no systematic differences from those we studied, but following up
with systematic study in these counties would be a good topic for future research.

18

 5

Verification by Direct Survey

We now verify the accuracy of our extrapolation in Section 4 to predicted 50c party members across China. To do this, we take the unusual step, in this context, of conducting a
sample survey of predicted 50c party members, along with elements designed to validate
this method of validation.5
1. Design

We began by creating a large number of pseudonymous social media ac-

counts. This required many research assistants and volunteers, having a presence on the
ground in China at many locations across the country, among many other logistically difficult complications. We conducted the survey via “direct messaging” on Sina Weibo,
which enables private communication from one account to another. With IRB permission, we do not identify ourselves as researchers and instead pose, like our respondents,
as ordinary citizens. Since information in our archive appears to indicate that government
monitoring of 50c party member activities occurs only through voluntary self-reporting
up the chain of command, our survey questions and the responses are effectively anonymous, which are conditions that have been shown to make respondents more sincere in
responding to sensitive questions (Tourangeau, Conrad and Couper, 2013).
We drew a random sample of social media accounts we predicted in Section 4 to be
50c and asked each whether he or she was indeed a 50c party member (in a special manner
described below). Of course, the difficulties of interpreting these answers is complicated
by the fact that our survey respondents are conducting surreptitious operations on behalf
of the Chinese government designed to fool users of social media into thinking that they
are ordinary citizens, and we are asking them about this very activity. In most cases, the
government is also their employer, and so they have ample incentives to not comply with
our requests, or to not comply sincerely.
We address these uncertainties with two entire additional surveys designed to provide
5

We had full IRB approval for our study (although we cannot make public our exact question wording
in Chinese, sample size, or data). We also added our own additional ethics rules not required by the IRB
designed to further protect the identities of our respondents and to keep our large research team safe. Our
rules followed the principle articulated in King, Pan and Roberts (2014, fn.20) of trying to avoid influencing
the system we were studying, which has the added advantage of reducing the chance for bias.

19

 internal checks on our results, as well as a carefully worded survey question in our anonymous survey context. In most surveys, researchers are left trusting the answer, perhaps
after a stage of pretesting or cognitive debriefing. In our survey, we are in the unusual position of being able to go further by offering a gold standard validation, where, for some
respondents, we know the outcome to question we are asking. That is, we ask the same
question of a random sample of known 50c party members from our Zhanggong leaked
archive. If the results of our survey of predicted 50c party members give similar results as
this survey, then we should have more confidence in the results.
We also offer a third entire survey that approximates the opposite gold standard by
asking those known not to be 50c party members. To do this, we draw a random sample
from Weibo accounts across China among those who do not engage with government
Weibo accounts and have more than 10 followers. Our results would be confirmed if
the percent who say they are 50c in this sample are significantly lower than those who
acknowledge being 50c in our predicted 50c sample. A tiny fraction of these accounts may
actually be 50c, but that would merely bias the results against the test of our hypothesis.
The final way we reduce uncertainty is in the design of our survey question. We
followed best practices in designing survey questions about sensitive topics, including
adjusting the perceived social environment (N¨aher and Krumpal, 2012) and using familiar
language and positive “loading” of sensitive questions (Groves et al., 2011). We also
studied a large volume of social media interactions, both via automated means (King, Lam
and Roberts, 2016) and by direct reading, and found a way within the cultural context
to ask the question so that it would be more likely to elicit a sincere answer. We also
pre-tested our survey on an independent sample. Although preserving the confidentiality
of our respondents and research team makes us unable to share the exact text of our
question, we report here a similar version in English, which will also enable us to explain
its features:
I saw your comment, it’s really inspiring, I want to ask, do you have any
public opinion guidance management, or online commenting experience?
To avoid interfering or influencing the system we are studying, and to avoid putting our
20

 respondent in an uncomfortable position, the question discusses online propaganda in
positive terms. We use the term “opinion guidance management” and “online commenting,” which we learned from our leaked archive is the way the government discusses these
tasks. We avoid terms like 50c or “Water Army” which has negative connotations for
some. Instead of asking someone to “out” themselves as a 50c party member, we ask for
advice on where the person learned to write in such a motivating, inspiring manner, thus
avoiding generating defensiveness on part of the respondent.
2. Results

High quality web surveys have response rates of about 3.5% (Pew Research

Center, 2014). The response rate for our survey was 6.5% which, although small on an
absolute level, is encouraging given our more challenging environment. In addition, unlike most web surveys, we were able to perform some checks for selection bias because
we collected available information on our entire target sample, before administering our
survey question. These variables for each social media account include the number of
followers, gender, year of creation, average number of posts for each month, and enabling geo-location; we also observed each of these variables within the five separate data
sources in Section 4. The vast majority of tests we conducted indicated statistically insignificant differences between respondents and nonrespondents. The few differences that
appeared are negligible compared to the large effect sizes we present below. As might be
expected, the data contains some evidence that 50c party members are less likely to respond to our question than non-50c party members, which has the effect of making it more
difficult to confirm our hypothesis. Although we could weight the results below by the
differences we found, they are small enough that we chose to present the raw, unprocessed
data below instead.
The results for our three surveys appear in Table 1. Overall, we found that 59% of our
predicted 50c party members admitted to being 50c party members. If we are correct that
they are all 50c party members, then the remaining 41% gave an insincere answer, which
would not be surprising given that this is essentially their job. To test this, we use our
sample of known 50c party members revealed in our leaked Zhanggong email archive. In
this sample, 57% admitted to their 50c party status. The two percentage point difference
21

 between these two figures is not statistically significant (at α = 0.05), suggesting that
indeed all respondents in our predicted sample are 50c. Also as a test, we use our sample
that approximates those known to not be 50c party members. In this sample, only 19%
said they were 50c; the 40 percentage point difference between this figure and that from
our predicted 50c party member sample (59%) is statistically significant, revealing that
we have detected a strong signal of actual 50c party membership among our predicted
50c sample.
Overall, the results from this survey strongly support the validity of the predictions of
50c party membership conducted in Section 4.
50c Status
Origin
Percent Yes
Predicted 50c
Across China
59%
Known 50c
Leaked Zhanggong Archive
57%
Known “Not” 50c Random sample
19%
Table 1: Survey About 50c Status. The first line is our survey; the second two are gold
standard evaluation surveys. The difference between the first and second lines is not statistically significant; the difference between the first and the third is statistically significant
(both at α = 0.05).

6

Size of the 50c Party

In this section, we study how widespread 50c activity is. Overall, we find a massive
government effort, where every year the 50c party writes approximately 448 million social
media posts nationwide. About 52.7% of these posts appear on government sites. The
remaining 212 million posts are inserted into the stream of approximately 80 billion total
posts on commercial social media sites, all in real time. If these estimates are correct,
a large proportion of government web site comments, and about one of every 178 social
media posts on commercial sites, are fabricated by the government. The posts are not
randomly distributed but, as we show in Figure 2, are highly focused and directed, all
with specific intent and content. The rest of this section explains how we estimate these
numbers. Throughout, in lieu of the possibility of formal standard error calculations, we
offer transparent assumptions that others can easily adjust to check sensitivity or improve
as more information is unearthed.
22

 1. Number of Social Media Posts To understand the context into which 50c posts are
inserted, we begin by estimating the total number of Chinese social media posts nationwide. As of December 2012, netizens were posting approximately 100 million messages a
day, or 36.5 billion a year, on Sina Weibo alone (Zhao et al., 2014), which is one of at least
1,382 known social media sites (King, Pan and Roberts, 2013). In our data, the ratio of
Sina Weibo posts to all posts is 1.85, meaning that an estimate of the total number of posts
on all platforms is (1.85 × 36.5 billion =) 67.5 billion. However, this requires the strong
assumption that 50c party members use specific commercial social media platforms in the
same proportions as the entire user population. We therefore use the detailed survey from
iiMedia Research Group (2014) and calculate that the ratio of total posts to Sina Weibo
posts to be 2.10, and the total number of posts per year to be about 80.4 billion. This is
an underestimate because it is based on microblogs and ignores blogs, but blogs probably
number in the millions which is rounding error on this scale.
2. Number of 50c Posts in Zhanggong

Among the 43,797 confirmed 50c posts, 30,215

were made during a 365 day period between 2/11/2013 (the first day on which we observed a 50c post) and 2/10/2014. We have evidence of at least 1,031 exclusive (Sina
Weibo) accounts in Zhanggong, including 202 accounts in the leaked archive and 829 we
identified outside the archive (by following the rules in Section 4).
In our archive, a 50c party member needing to make a post chooses an exclusive account on Weibo (689/43, 797 =) 1.57% of the time compared to all other choices (an
ordinary account on Weibo or another social media site). We assume this ratio is approximately the same for non-leaked 50c posts in Zhanggong, which in turn implies that the
ratio of total 50c posts to 50c posts in the archive is the same as the ratio of total exclusive
accounts to exclusive accounts in the archive. As such, an estimate of the total number of
posts in Zhanggong in 2013 is (30, 215 × 1, 031/202 =) 154,216.
3. Number of 50c Posts in Jiangxi Province Zhanggong is an urban district of Ganzhou
City within Jiangxi province. According to China Internet Network Information Center’s
Statistical Report on Internet Development in China 2014, the 2013 Internet penetration of

23

 urban residents was 62.0% and of rural residents was 27.5% (http://j.mp/CNNIC).
According to the National Bureau of Statistics of China, 48.87% of the 45.22 million people in Jiangxi province lived in urban areas, or 22.10 million, with 23.12 million living in
rural areas (http://j.mp/ChinaSY).
We first compute the number of 50c posts per internet user in Zhanggong, which is
(154, 216/468, 461 × 0.62 =) 0.531. We then assume that this rate is roughly the same in
Jiangxi and then scale up. Thus, we estimate a total number of 50c posts in Jiangxi during
2013 as (0.531 × [0.62 × 22.1M + 0.275 × 23.1M ] =) 10.65 million.
4. Number of 50c Posts in China Finally, to scale this result to all of China, we assume
that the ratio of 50c posts to population in other parts of China is roughly the same as in
Jiangxi. This ratio of posts per users is (10.65M/14.68M =) 0.7255. Applying this
assumption to the country as a whole reveals the presence of (0.7255 × 617.58 =) 448.0
million 50c posts in China during 2013.

7

Theoretical Implications

The empirical results offered above seem clear, but what do they suggest about the overall
strategy of the Chinese government, or for authoritarian regimes in general? We first explain these results by generalizing prior findings on (human) censorship and (automated)
filtering, all led by the same propaganda department in the same government as the 50c
party (King, Pan and Roberts, 2013, 2014). We then extend these ideas to the authoritarian
literature in general.
1. China One way to parsimoniously summarize existing empirical results about information control in China is with a theory of the strategy of the regime. This theory involves
two simple and complementary principles the regime appears to follow, one passive and
one active. The passive principle is do not engage on controversial issues, including no
50c posts supporting, and no censoring of posts criticizing, the regime, its leaders, or their
policies.

24

 To develop the active principle, we first define social media posts with “collective action potential” as those about real world crowd formation and related activities (see King,
Pan and Roberts, 2013, p.6, for a more precise definition). Citizens criticize the regime
without collective action on the ground in many ways, including even via (unsubstantiated) threats of protest and viral bursts of online-only activity — all which do not have
collective action potential by this definition and so are ignored by the government. Then
the second, active, principle is stop discussions with collective action potential, by active
distraction and active censorship. Cheerleading in directed 50c bursts is one way the government distracts the public, although this activity can be also be used to distract from
other events, general negativity, specific grievances, etc.6
These principles appear to derive from the fact that the main threat perceived by the
Chinese regime in the modern era is not military attacks from foreign enemies but rather
uprisings from their own people. Staying in power involves managing their government
and party agents in China’s 32 provincial-level regions, 334 prefecture-level divisions,
2,862 county-level divisions, 41,034 township-level administrations, and 704,382 villagelevel subdivisions, and somehow keeping in check collective action organized by those
outside of government. The balance of supportive and critical commentary on social media about specific issues, in specific jurisdictions, is useful to the government in judging
the performance of (as well as keeping or replacing) local leaders and ameliorating other
information problems faced by central authorities (Dimitrov, 2014a,b,c; Wintrobe, 1998).
As such, avoiding any artificial change in that balance — such as from 50c posts or censorship — can be valuable.
Distraction is a clever and useful strategy in information control in that an argument
in almost any human discussion is rarely an effective way to put an end to an opposing
argument. Letting an argument die, or changing the subject, usually works much better
than picking an argument and getting someone’s back up (as new parents recognize fast).
It may even be the case that the function of reasoning in human beings is fundamentally
about winning arguments rather than resolving them by seeking truth (Mercier and Sper6

Unfortunately, although our leaked archive includes directions to 50c workers, the information does not
reveal whether these directions originate from Zhanggong or from higher levels of the government or party.
This prevents us from determining with any certainty the specific intent of each burst of 50c activity.

25

 ber, 2011). Distraction even had the advantage of reducing anger compared to ruminating
on the same issue Denson, Moulds and Grisham (2012). Finally, since censorship alone
seems to anger people Roberts (2015), the 50c astroturfing program has the additional advantage of enabling the government to actively control opinion without having to censor
as much as they might otherwise.
Our inference about distraction being the goal of the regime is consistent with directions to 50c party members in emails from the Zhanggong propaganda department. They
ask the 50c members to “promote unity and stability through positive publicity” (坚持“
团结稳定 鼓劲、正面宣传为主) and “actively guide public opinion through emergency
events” (积极稳妥做好突发事件舆论引导). In this context, “emergency events” are
events with collective action potential.7
2. Authoritarian Politics For the literature on authoritarian politics in general, our results may help refine current theories of the role of information, and in particular what is
known as “common knowledge,” in revolutionary mobilization. Many theories in comparative politics assume that autocrats slow the spread of information critical of the regime in
order to minimize the development of common knowledge of grievances which, in turn,
reduces the probability of mobilization against the regime. The idea is that coordination
is essential to revolution, and coordination requires some common knowledge of shared
grievances (Chwe, 2013; Egorov, Guriev and Sonin, 2009; Hollyer, Rosendorff and Vreeland, 2014; Persson and Tabellini, 2006; Tilly, 1978).
In contrast, our results seem to indicate that the Chinese regime differentiates between
two types of common knowledge — about specific grievances, which they allow, and
7

For example, a website developed by the Ministry of Public Security and Ministry of Education to
help young people better understand safety issues (http://j.mp/EmergEvents) explains: “Every
emergency event involves the self interest of a particular group of people, leading to psychological pressure
and change among this group, and understandably leading to concern and worry. Especially for emergency
events of a societal nature [as distinct from natural disasters], most are organized by a small group of people,
who through their publicity-seeking and encouragement get others involved. Recently, emergency events
due to issues like territorial disputes, land requisitions, and housing demolition in certain areas are often
organized by one person and involve many, making collective events.” (任何一类突发事件， 都必然要
涉及一部分人的切身利益，使其产生心理压力和变化，引起人们的 关注和不安也属正常。尤其是
社会性的突发事件，多是由少数人操纵，通过 宣传鼓动把一些群众卷到事件中来。近期，在一些
地方因地界、征地、拆迁 安置而发生的突发性事件，往往是一人纠合，数人响应，使其具有聚众
性。.)

26

 about posts with collective action potential, which they do a great deal to avoid. Avoiding
the spread of common knowledge about collective action events (and not grievances) is
consistent with research by Kuran (1989, 1991), Lohmann (1994), and Lorentzen (2010),
who focus specifically on the spread of information about real-world protest and on-going
collective action rather than the generic spread of common knowledge more broadly.
The idea is that numerous grievances of a population ruled autocratically by nonelected leaders are obvious and omnipresent. Learning of one more grievance, in and
of itself, should have little impact on the power of a potential revolutionary to ignite
protest. The issue, then, appears not to be whether such grievances are learned by large
enough numbers to foment a revolution. Instead, we can think of creative political actors,
including those aspiring to lead a revolution or coup, as treating issues, ideologies, events,
arguments, ideas, and grievances as “hooks on which politicians hang their objectives and
by which they further their interests,” including interests that entail initiating or fostering
a political uprising (Shepsle, 1985). If one hook isn’t available, they can use another.
By this logic, then, common knowledge of grievances is already commonplace and so
allowing more information about them to become public is of little risk to the regime or
value to its opponents. Since disrupting discussion of grievances only limits information
that is otherwise useful to the regime, the leaders have little reason to censor it, argue with
it, or flood the net with opposing viewpoints. What is risky for the regime, and therefore
vigorously opposed through large scale censorship and huge numbers of fabricated social
media posts, is posts with collective action potential.

8

Concluding Remarks

The vast majority of scholars, journalists, activists, and participants in social media have,
until now, been convinced that the massive 50c party is devoted to engaging in argument
that defends the regime, its leaders, and their policies. Our evidence indicates the opposite
— that the 50c party engages in almost no argument of any kind and is instead devoted
primarily to cheerleading for the state, symbols of the regime, or the revolutionary history
of the Communist Party. We interpret these activities as the regime’s effort at strategic

27

 distraction from collective action, grievances, or general negativity, etc.
It also appears that the 50c party is mostly composed of government employees contributing part time outside their regular jobs, not, as has been claimed, ordinary citizens
paid piecemeal for their work. This, nevertheless, is still an enormous workforce that, we
estimate, produces 488 million 50c posts per year. Their effectiveness appears maximized
by the effort we found of them concentrating the posts into spikes at appropriate times,
and by directing about half of the posts to comments on government web sites.
Inferences in this paper depend on the veracity of the leaked archive we analyze. The
size and extraordinary complexity of this archive makes it highly likely to be real. There
are no signs of it having been generated by automated means, and doing so by hand
would have been a monumental task. We have also verified numerous external references
from the data — to specific individuals, email addresses, phone numbers, government
departments, programs, web sites, social media accounts, specific posts, etc. — and every
one checks out.
Of course, 50c party efforts may exist that do not follow the model we unearthed in
Zhanggong, and it is possible that such hypothetical organizations may follow different
rules and practices, perhaps in different places, and of course may generate 50c posts
with unrelated content. We think this is unlikely, but regardless, perhaps the window this
paper opens on this enormous and previously secretive government program may help
others discover different aspects of it in China, and eventually in other related authoritarian regimes.

A

Appendix: Categorization Scheme

Our categorization scheme for social media posts includes the six categories below, along
with examples of each. Non-Chinese speakers should be aware when reading these examples that the Chinese language, even on social media, tends to be quite flowery and
formal, with frequent creative, and often (to English speakers) stagy-sounding, wordings.
(1) Taunting of Foreign Countries Favorable comparisons of China to other countries; includes insults to other countries, and taunting of pro-democracy, pro-West, proindividual liberties, or pro-capitalist opinions within China. Examples from leaked Zhanggong 50c posts:
28

 • 去年，奥巴马在香格里拉会议上力邀23国参与围堵中国时这样说 道：“中国
有13亿人，他们越崛起，我们就会越没饭吃，因为地球资源供给 是有上限
的。所以为了我们能继续过现在的生活，就必须遏制中国的发 展。” [Last
year, at the Shangri-la Dialogue where Obama invited 23 countries to participate in
the containment of China, he said: “China has 1.3 billion people, the faster China
rises, the more difficult it will be for us to live, because the earth’s resources are
limited. For us to remain at our current living standard, we must contain China’s
development.”]
• 中国的崛起大势已经不可阻挡。美国一边公开宣称不是中国死就 是西方
亡，一边又拼命告诉中国民众：你们的政府有问题啊，必须推翻它， 然后
你们就能过上比现在更好的日子。——请问，还有比这更可笑和自相矛盾
的逻辑吗？ [China’s rise is now inevitable. On one hand, the US publicly asserts
that if China does not perish the West will wither; on the other hand it tells the
Chinese people that your government is problematic: you have to overthrow it so
you can live a better life. Is there a more ridiculous and contradictory logic than
this?]
(2) Argumentative praise or criticism Comments on controversial, Pro/Con (nonvalience) issues, as well as claims of wrongdoing or unfairness, praise (usually of the
government) or criticism (usually of opponents of the government), taking a position or
explaining why a particular viewpoint is correct or (more often) wrong. These posts are
often part of a debate, in opposition to a previous post. Examples from leaked Zhanggong
50c posts:
• 我亲爱的朋友们，翻一下你的微薄，你会发现系统已经自动帮你添加了 诸
如薛蛮子、李开复、作业本、韩寒、李承鹏等各种民粹微薄，这是标准
的强制灌输和 洗脑手段，建议你取消关注 [My dear friends, you if you go
through your Weibo, you’ll discover that the system automatically had you follow
Xue Manzi, Li Kaifu, Zuo Yeben, Han Han, Li Chengpeng and other populist Weibo
users. This is a typical tactic of indoctrination and brainwashing, I suggest you
unfollow them.]
• 李开复说纽约60万美金一套别墅，比北京便宜多了，但他不会告 诉你那套
所谓的别墅其实是个仓库，而且离纽约市区开车需要四个多小时 [Li Kaifu
says that you can buy a villa for $600,000 USD in New York, much cheaper than
in Beijing. But what he doesn’t tell you is that this so-called villa is actually a
warehouse, which is more than a four hour drive from New York City.]
(3) Non-argumentative Praise or Suggestions Focused on noncontroversial valience
issues which no one could argue against, such as improving housing and public welfare. It
includes praise of current government officials, programs, or policies. It doesn’t respond
to alternative, opposing viewpoints, and it includes positive sentiment. It is distinguished
from category (3) in that it praises something specific such as the government, its officials,
government programs, or initiatives. It includes a small number of constructive suggestions for what government policies might include (i.e., added benefits rather than critical
complaints). It does not argue against a specific viewpoint, but just says “it would be
29

 nice if the government did X,” which usually the government is already in the process of
implementing. Some examples of known Zhanggong 50c posts:
• 政府. . . 做了好多实事，其中解决了好大一部分的人住房。 [The government
has done a lot of practical things, among which is solving a significant part of the
housing problem]
• 土坯房改造政策，让我们村的人搬出了坯房都住上了小洋楼， 村 里发生了
翻天覆地的变化，真是太感谢了。 [The policy of renovating mud-brick houses
has allowed villagers to move out of mud-brick dwellings into small, Western-style
buildings. The village has been transformed, we are so grateful]
• 希望中央支持力度更大 [We hope the central government provides us with even
more support]
• 希望能有更多类似《若干意见》的好政策！ [We hope there will be more good
policies like ”Various Opinions” (the abbreviated name of an economic development policy)]
• 期待书记带领我们. . . 特别是在教育、医疗卫生方面，争取更 多政策为百姓
谋取更大福祉。 [We look forward to the leadership of our party secretary. . . We
hope that he can carry out more policies that will benefit the people in different
aspects, especially in education and health care.]
(4) Factual Reporting Descriptions of current government programs, projects, events,
or initiatives, or planned or in progress initiatives. Does not include any praise of these
programs or events (which would be category (3)), just that they are occurring. Reporting
on what government, government officials are doing. Some examples of known Zhanggong 50c posts:
• 清明三天假期7座小客车继续免费 [During the Qingming festival three-day holiday, [the freeway] will remain free to 7-seater busses]
• 6月27日，江西省委作出向龚全珍同志学习的决定， 号召全省党员干部深
入学习龚全珍坚定信念、永葆本色的坚强党性， 服务群众、一心为民的
质朴情怀，忠于职守、矢志不渝的执着追求，淡泊名利、 无私奉献的高尚
情操，勤俭节约、艰苦朴素的生活作风。 [On June 27, the Jiangxi provincial
committee promulgated an opinion to learn from comrade Zhen Gongquan, calling
on all provincial party members and cadres to study Zhen Gongquan’s firm conviction, staunch support of the Party’s spirit, service to the masses, straightforward
dedication to the people, devotion to duty, abiding dedication, indifference to fame
and fortune, selfless dedication to moral character and hardwork.]
• 1月16日，江西省委常委、赣州市委书记史文清将通过中国赣州网与 网民在
线交流，倾听网民意见、建议和诉求。 [On January 16, Jiangxi Party Committee Member and Ganzhou City Party Secretary Shi Wenqing will communicate with
netizens on the China Ganzhou Web, to hear comments, suggestions, and demands
from netizens.]

30

 (5) Cheerleading for China Patriotism, encouragement and motivation, inspirational
quotes and slogans, inspirational quotes from government officials, thankfulness, gratefulness, inspiration or thankfulness for historical and aspirational figures or events, and
cultural references and celebrations (e.g., describes traditions, actions, suggestions for the
community). Does not include positive sentiment toward particular government leaders
or specific policies, but does include positive sentiment or general praise toward life, historical figures, model citizens (e.g., Lei Feng; Gong Quanzhen, a model teacher; Guo
Chuhui, a patriotic villager), or China in general. Some examples of known Zhanggong
50c posts:
• 众多革命先烈们的英勇奋斗，缔造了我们今天的幸福生活！ 向英雄致敬。
[Many revolutionary martyrs fought bravely to create the blessed life we have today!
Respect these heroes.]
• 向所有为中华民族繁荣富强做出伟大贡献的先人们致敬！人民英雄永垂不
朽 [Respect to all the people who have greatly contributed to the prosperity and
success of the Chinese civilization! The heroes of the people are immortal]
• 接过父辈、祖辈血染的红旗，坚定不移地跟着党走！ [[I will] carry the red
flag stained with the blood of our forefathers, and unswervingly follow the path of
the CCP!]
• 我们自己要更加努力，不等不靠，主动上前。 [We all have to work harder, to
rely on ourselves, and to take the initiative to move forward.]
• 爱我中华 [I love China]
• 大家的日子都过好了，中国梦就实现了！ [[If] everyone can live good lives,
then the China Dream will be realized!]
• 赣州加油哦 [Way to go Ganzhou]
(6) Other Irrelevant posts that are entirely personal, commercial (such as ads), jokes,
or empty posts that forward information not included. This category is removed and
conditioned on in all analyses in the paper.

References
Bandurski, David. 2008. “China’s guerrilla war for the web.” Far Eastern Economic
Review 171(6):41.
Barr, Michael. 2012. “Nation branding as nation building: China’s image campaign.” East
Asia 29(1):81–94.
Bremmer, Ian. 2010. “Democracy in Cyberspace.” Foreign Affairs 89(6):86–92.
Chen, Jidong, Jennifer Pan and Yiqing Xu. 2016. “Sources of Authoritarian Responsiveness: A Field Experiment in China.” American Journal of Political Science 60(2):383–
400.
Chwe, Michael Suk-Young. 2013. Rational ritual: Culture, coordination, and common
knowledge. Princeton University Press.
31

 Deibert, Ronald and Rafal Rohozinski. 2010. “Liberation vs. control: The future of cyberspace.” Journal of Democracy 21(4):43–57.
Denson, Thomas F, Michelle L Moulds and Jessica R Grisham. 2012. “The effects of analytical rumination, reappraisal, and distraction on anger experience.” Behavior therapy
43(2):355–364.
Dimitrov, Martin K. 2014a. “Internal Government Assessments of the Quality of Governance in China.” Studies in Comparative International Development pp. 1–23.
Dimitrov, Martin K. 2014b. “Tracking Public Opinion Under Authoritarianism.” russian
history 41(3):329–353.
Dimitrov, Martin K. 2014c. “What the Party Wanted to Know Citizen Complaints as
a “Barometer of Public Opinion” in Communist Bulgaria.” East European Politics &
Societies 28(2):271–295.
Editors. 2013. “Cat and mouse: How China makes sure its internet abides by the rules.”
The Economist April 6.
Editors. 2016. “Fifty cents.” China Digital Times . http://j.mp/cdt50cent.
Egorov, Georgy, Sergei Guriev and Konstantin Sonin. 2009. “Why Resource-poor Dictators Allow Freer Media: A Theory and Evidence from Panel Data.” American Political
Science Review 103(4):645–668.
Freedom House. 2009. “China and the Internet: An Uphill Fight for Freedom.” Harvard
International Review pp. 68–73.
Greitens, Sheena Chestnut. 2013. “Authoritarianism Online: What Can We Learn from
Internet Data in Nondemocracies?” PS: Political Science & Politics 46(02):262–270.
Groves, Robert M, Floyd J Fowler Jr, Mick P Couper, James M Lepkowski, Eleanor
Singer and Roger Tourangeau. 2011. Survey methodology. Vol. 561 John Wiley &
Sons.
Haley, Usha. 2010. “China’s Fifty Cent Party for Internet Propaganda.” Huffington Post
October 4.
Han, Rongbin. 2015. “Manufacturing Consent in Cyberspace: China’s “Fifty-Cent
Army”.” Journal of Current Chinese Affairs 44(2):105–134.
Hassid, Jonathan. 2012. “Safety valve or pressure cooker? Blogs in Chinese political
life.” Journal of Communication 62(2):212–230.
Henochowicz, Anne. 2014. “Thousands of Local Internet Propaganda Emails Leaked.”
China Digital Times . http://j.mp/leakedEmails.
Hollyer, James R., B. Peter Rosendorff and James R. Vreeland. 2014. “Measuring transparency.” Political analysis 22(4):413–434.
Hopkins, Daniel and Gary King. 2010. “Improving Anchoring Vignettes: Designing
Surveys to Correct Interpersonal Incomparability.” Public Opinion Quarterly pp. 1–22.
http://gking.harvard.edu/files/abs/implement-abs.shtml.
iiMedia Research Group. 2014. “2013 年中国移动社会化分享年度报告.” China Mobile
Social Share Annual Report (内 部版) . http://j.mp/iimedia.
Keating, Joshua. 2011. “China’s ‘50-cent’ party takes on the jasmine revolutions.” Foreign
Policy March.
King, Gary, Jennifer Pan and Margaret E. Roberts. 2013. “How Censorship in China
Allows Government Criticism but Silences Collective Expression.” American Political
Science Review 107:1–18. http://j.mp/LdVXqN.
King, Gary, Jennifer Pan and Margaret E. Roberts. 2014. “Reverse-engineering cen32

 sorship in China: Randomized experimentation and participant observation.” Science
345(6199):1–10. http://j.mp/1KbwkJJ.
King, Gary, Patrick Lam and Margaret Roberts. 2016. “Computer-Assisted Keyword and
Document Set Discovery from Unstructured Text.”.
Kuran, Timur. 1989. “Sparks and prairie fires: A theory of unanticipated political revolution.” Public Choice 61(1):41–74.
Kuran, Timur. 1991. “Now out of never: The element of surprise in the East European
revolution of 1989.” World politics 44(1):7–48.
Lam, Oiwan. 2012. “China: Challenging the 50 Cent Party.” Global Voices March.
Lam, Oiwan. 2013. “China Beefs Up ‘50 Cent’ Army of Paid Internet Propagandists.”
Global Voices October.
Letham, Benjamin, Cynthia Rudin, Tyler H McCormick, David Madigan et al. 2015.
“Interpretable classifiers using rules and Bayesian analysis: Building a better stroke
prediction model.” The Annals of Applied Statistics 9(3):1350–1371.
Lohmann, Susanne. 1994. “The Dynamics of Informational Cascades: The Monday
Demonstrations in Leipzig, East Germany, 1989-1991.” World Politics 47(1):42–101.
Lorentzen, Peter. 2010. “Regularizing Rioting: Permitting Protest in an Authoritarian
Regime.” Working Paper.
Mercier, Hugo and Dan Sperber. 2011. “Why do humans reason? Arguments for an
argumentative theory.” Behavioral and brain sciences 34(02):57–74.
Miller, Blake Andrew Phillip. 2016. “Automatic Detection of Comment Propaganda in
Chinese Media.” Available at SSRN 2738325 .
N¨aher, Anatol-Fiete and Ivar Krumpal. 2012. “Asking sensitive questions: the impact of
forgiving wording and question context on social desirability bias.” Quality & Quantity
46(5):1601–1616.
Ng, Jason. 2011. “Wu Mao.” ‘Blocked on Weibo’ blog . blockedonweibo.tumblr.com.
Persson, Torsten and Guido Tabellini. 2006. “Democracy and development: The devil in
the details.” American Economic Review 96:319–324.
Pew Research Center. 2014.
“Beyond Red vs. Blue: The Political Typology.”
http://j.mp/PewResponse.
Philipp, Joshua. 2015. “Leaked Emails Show Chinese Regime Employs 500,000 Internet
Trolls.” Epoch Times . http://j.mp/twomill.
Roberts, M. 2015.
“Experiencing Censorship Emboldens Internet Users and
Decreases Government Support in China.” Unpublished Working Paper, URL
http://www.margaretroberts.net/wp-content/uploads/2015/07/fear.pdf .
Shepsle, Kenneth A. 1985. Comment of Why the Regulators Chose to Deregulate. In
Regulatory Policy and the Social Sciences, ed. Roger Noll. Berkeley: University of
California Press pp. 231–239.
Sonnad, Nikhil. 2014. “Hacked emails reveal China’s elaborate and absurd internet propaganda machine.” Quartz . http://j.mp/Sonnad.
Strafella, Giorgio and Daria Berg. 2015. ““Twitter Bodhisattva”: Ai Weiwei’s Media
Politics.” Asian Studies Review 39(1):138–157.
Tilly, Charles. 1978. From mobilization to revolution. McGraw-Hill New York.
Tong, Yanqi and Shaohua Lei. 2013. “War of position and microblogging in China.”
Journal of Contemporary China 22(80):292–311.
Tourangeau, Roger, Frederick Conrad and Mick Couper. 2013. The science of web sur33

 veys. Oxford University Press.
Wintrobe, Ronald. 1998. The Political Economy of Dictatorship. Cambridge University
Press.
Yang, Xiaofeng, Qian Yang and Christo Wilson. 2015. Penny for Your Thoughts: Searching for the 50 Cent Party on Sina Weibo. In Ninth International AAAI Conference on
Web and Social Media.
Zhao, Juanjuan, Weili Wu, Xiaolong Zhang, Yan Qiang, Tao Liu and Lidong Wu. 2014.
“A short-term trend prediction model of topic over Sina Weibo dataset.” Journal of
Combinatorial Optimization 28(3):613–625.

34