Fighting Propaganda with Censorship: A Study

of the Ukrainian Ban on Russian Social Media

Yevgeniy Golovchenko, University of Copenhagen

Many states have become concerned with Russian cyberattacks and online propaganda. The Ukrainian government re-
sponded to the information threat in 2017 by blocking access to several Russian websites, including VKontakte, one of the
most popular social media websites in Ukraine. By exploiting a natural experiment in Ukraine, I find that the sudden cen-
sorship policy reduced activity on VKontakte, despite the fact that a vast majority of the users were legally and technically able
to bypass the ban. Users with strong political and social affiliations to Russia were at least as likely to be affected by the ban as
those with weak affiliations. I argue that the ease of access to online media—not political attitudes toward the state—was the
main mechanism behind the users’ response to the ban. These findings suggest that this pragmatic view on the effects of
censorship holds, even in the highly politicized military conflict between Russia and Ukraine, which later evolved into a full-
scale Russian invasion.

T o what extent does censorship reduce activity on banned

media among different societal groups in the context of
international conflict? The answer to this question is es-
sential if one is to understand the factors enabling or limiting the
capacity of the state to control the flow of online information.
to major Russian websites, including VKontakte, the second
most visited social media platform in Ukraine (following You-
Tube) at the time (Roth 2017). The Kremlin’s control over Rus-
sian social media was one of the reasons why the Ukrainian
government viewed the VKontakte ban as a national security
Authoritarian and nonauthoritarian states alike use cen- measure against Russian propaganda and surveillance.
sorship to police cyberspace (Edwards 2009; Howard, Agarwal, How effective are such interventions? Scholarly literature,
and Hussain 2011). In recent years, an increasing number of ranging from media studies to political science, suggests that
European states and tech firms have used it to combat digital censorship may successfully limit the overall access to infor-
disinformation and foreign interference. These information mation (Kalathil and Boas 2010; Morozov 2012; Roberts 2018;
threats are often attributed to the Russian government and Zhang et al. 2017). Censorship can also backfire via the so-
portrayed by the authorities as a threat to national security and called Streisand effect, by drawing attention to the forbidden
election integrity (BBC 2017; Fiorentino 2018). In this context, information, causing political outrage against the censor, and
Ukraine offers some of the most extreme examples of cen- increasing public awareness or interest in the forbidden content
sorship use as a response to information war and online pro- (Jansen and Martin 2015). In addition to this, bans may in-
paganda from abroad. centivize users to master circumvention through virtual private
In 2017, the Ukrainian government issued an executive network (VPN) services, thereby gaining access to even more
order, which forced internet service providers to block access forbidden platforms (Hobbs and Roberts 2018). Conversely,

Important disclaimer: This article was written prior to the Russian Federation's full-scale invasion of Ukraine on February 24, 2022, which is why it mainly
reflects events prior to this. Edits have been made to make the reader aware of the invasion. However, the changes are minimal in order to stay true to the
accepted version.
Yevgeniy Golovchenko (yg@ifs.ku.dk) is a postdoc in the Department of Political Science, University of Copenhagen.
This work is part of the Digital Disinformation and DIPLOFACE research projects, both led by Rebecca Adler-Nissen. The former project (number CF16-
0012) is funded by the Carlsberg Foundation, while the latter project (number 680102) is funded by the European Research Council. The use of data was
approved by the University of Copenhagen. This project does not publish the data about VKontakte users because of the website’s terms of service. The code
used for the analysis and data collection from VKontakte’s public application programming interface and replication files are available in the JOP Dataverse (https://
dataverse.harvard.edu/dataverse/jop). The empirical analysis has been successfully replicated by the JOP replication analyst. An online appendix with supplementary
material is available at https://doi.org/10.1086/716949.

Published online March 11, 2022.

The Journal of Politics, volume 84, number 2, April 2022. q 2022 Southern Political Science Association. All rights reserved. Published by The University of
Chicago Press for the Southern Political Science Association. https://doi.org/10.1086/716949 639
640 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

the decision not to censor ideologically “undesired” content or influence have reduced the wall posting activity on VKontakte
the inability to fully implement a ban can in some cases help among users with pro-Russian attitudes at least as much as
stabilize authoritarian regimes by providing citizens with among pro-Ukrainian users. I find the same pattern when
highly demanded entertainment and news about societal prob- comparing citizens in Ukraine with few social ties to citizens
lems in liberal democracies (Kern and Hainmueller 2009). within Russia versus those who are strongly embedded in
In sum, the existing studies document diverging instances the Russian social network. The increasing costs of using the
when censorship backfires in some scenarios and successfully banned platform (in terms of time and effort) explain the re-
limits information in other cases. Similarly, the research on sponse to censorship much more than do the social ties or po-
censorship presents examples when censorship is met by pub- litical attitudes toward the states involved in the armed conflict.
lic outrage or discontent (Jansen and Martin 2015; Roberts It is important to note that the ban has increased the cost of
2018) by some groups and political support by others (Esberg going online to a relatively small extent when considering the
2020). Despite this, the existing literature offers little empirical accessibility of free VPN tools in Ukraine and the fact that
research regarding the extent to which the same censorship users are not legally persecuted for circumventing the ban.
policy may affect opposing societal groups differently (i.e., Even a small increase in the cost of accessing the Russian plat-
backfiring in one group while being effective in others). This form is enough to disrupt online activity among pro-Russian
question is crucial when evaluating the consequences of gov- (and pro-Ukrainian) users, who would rather shift to cheaper
ernment censorship. Even if a government succeeds in par- and more accessible alternatives. In other words, the accessi-
tially reducing the overall online activity on forbidden media, bility of the media appears to play a much more important role
the ban may backfire if the supporters of the regime become in the decision to use censored social media than do politics or
less active on the censored platform than the opposition. In social ties with citizens in the hostile state. The findings are in
other words, the government would risk making the opposi- line with the “accessibility view” on censorship. The theory
tion more prevalent on the platform than the supporters of the emphasizes the costs of accessing forbidden content—and not
regime. political attitudes—as the main mechanism behind the effect
The Ukrainian context highlights the importance of this of censorship on online behavior. This view has been previ-
question. If the VKontakte ban reduces social activity among ously used to explain the effects censorship among impatient
pro-Ukrainian users more than among pro-Russian users, the and relatively apolitical users in China through subtle “fric-
anti-Kremlin policy would make profiles with close ties to tion” in online access (e.g., by slowing down the connection;
Russia more prevalent in relative terms. From a surveillance Hobbs and Roberts 2018; Roberts 2018). This study suggests
point of view, this would give Russian authorities less access to that the relatively pragmatic view on censorship holds even in
data on pro-Ukrainian users (e.g., activists, Ukrainian volun- the highly politicized context of the hybrid Russia-Ukraine war
teers, or soldiers). From a propaganda point of view, however, (Lanoszka 2016; Reisinger and Gol’c 2014), which later turned
the censorship would backfire by leaving pro-Kremlin pro- into a Russian full-scale invasion of Ukraine.
paganda even less contested on VKontakte. Drawing on the This article contributes to the literature on censorship by
Ukraine case, this article seeks to answer the following ques- addressing the debates on the driving mechanisms behind
tion: What is the effect of the Ukrainian ban on online activity large-scale bans (Bunn 2015; Lorentzen 2014; Roberts 2018).
among VKontakte users with close affiliations to Ukraine and Furthermore, this research adds to the burgeoning literature
Russia? on misinformation (Nyhan and Reifler 2015; Tucker et al.
I approach the question empirically by using publicly avail- 2018) and propaganda (Peisakhin and Rozenas 2018; Slutsky
able data from VKontakte and a natural experiment research and Gavra 2017; Stukal et al. 2017), by empirically examining
design to estimate the causal effect of the ban on online activity the effects of one of the most radical, large-scale policies to
among different user groups. It is important to distinguish on- combat online manipulation.
line activity—measured as number of public posts uploaded by
the users on their own respective walls—from mere access to BACKGROUND
the VKontakte profiles (i.e., the ability to log on the platform). Russian authorities view the internet as a strategically impor-
The findings indicate that a vast majority of Ukrainians tant domain closely tied to national security (Vendil Pallin
on VKontakte were able to circumvent censorship by logging 2017). Following a series of legal restrictions in 2014, VKon-
back on to the forbidden website—likely through tools like takte and other Russian websites have been under increasing
VPN. Nevertheless, the Ukrainian government succeeded in pressure to hand over private information about its users to
reducing the overall online activity among Ukrainians on the FSB, the Russian Federal Security Service. This includes in-
Russian platform. Government attempts at curbing Russian formation about administrators behind Euromaidan-related
 Volume 84 Number 2 April 2022 / 641

pages (Sanovich 2017, 12). VKontakte’s founder, Pavel Durov, recruitment for the separatist movement (Deutsche Welle 2017).
initially attempted to resist the pressure from authorities (Pan In this sense, the VKontakte users are potentially exposed
2017). Eventually, he was pushed from the firm by major to pro-Kremlin content, through either their pro-Russian
shareholders and left Russia (Economic Times 2014). Today, friends and family or their newsfeed.
VKontakte is predominantly compliant with Russian author- The Russian surveillance “threat” is relevant for both ci-
ities, who have used the data in multiple criminal cases against vilian and military targets. Shklovski and Wulf (2018) find
individuals, some being prosecuted for antigovernment social that, despite military regulations, Ukrainian soldiers in the war
media posts (Interfaks 2018; Robinson 2018). zone use social media (including VKontakte) to search for in-
Ukrainian authorities responded to the Russian govern- formation and to maintain personal contacts, despite army
ment’s increasing influence over Russian social media in 2014 regulations. This even applies to soldiers who are aware that the
by advising citizens to delete their accounts on Russian-owned enemy may potentially use media surveillance to geolocate
social media websites (Boichak and Jackson 2020, 270) before Ukrainian positions and to select targets for artillery strikes (7, 10).
proceeding to a censorship policy in 2017. The VKontakte ban The VKontakte ban took place in a time of war but si-
is part of a decree that imposed sanctions on 468 organizations multaneously also in a relatively democratic context, where
and 1,228 individuals, including the Odnoklassniki social media citizens still enjoy legal access to a wide range of media outlets
platform and the Yandex search engine. VKontakte, however, and platforms. The ban was not intended to reduce the overall
was by far the most popular social media platform in Ukraine access to social media but rather to push users away from the
among all of the sites on the list. Russian VKontakte and toward other media alternatives, such
The goal behind the decree is “to protect the national se- as Facebook, Twitter, and Instagram, which are controlled by
curity and territorial integrity of Ukraine” (Dek, Kononova, neither the Russian nor the Ukrainian authorities.
and Marchenko 2019, 58). According to Oleksandr Turchynov,
secretary of the National Security and Defense Council of
THEORY: MECHANISMS BEHIND CENSORSHIP
Ukraine at the time of the ban, the forbidden websites were
The literature on censorship describes multiple factors that
used for Russian propaganda, state-driven surveillance, and
may cause people to abandon or desist from consuming, shar-
cyberattacks against Ukrainians (Andrusieczko 2017). The
ing, or producing forbidden content, as well as instances when
Ukrainian authorities argue that the ban is an important
censorship “backfires.” Overall, this literature can be divided
countermeasure against these information threats following the
into at least two strands, each with its own theoretical view on
Russian annexation of Crimea in 2014 and the information war
the mechanisms behind censorship. On one hand, censorship
on Ukraine (Kiselyova and Prentice 2017; RFE/RL 2017). As a
may work by increasing the “costs” of accessing forbidden
result of the annexation, the Ukrainian government could only
content in terms of time, money, or effort. I refer to this as the
fully implement the ban in territories under its control and not
“accessibility view” on censorship. On the other hand, cen-
in Crimea. For the purpose of this article, I refer to the affected
sorship can be seen as a signal that calls the citizens to fall in
region north of the Crimean peninsula as “Mainland.”
line with government policy, creating fears of reprisal or social
To this day, Ukraine remains divided by political tension
exclusion. For the sake of this article, I will refer to this line of
between pro-Ukrainian citizens (consisting of both ethnic
thought as the “political-signaling view.” As I argue below, the
Ukrainians and Russians), who support Ukrainian sover-
two strands explain not only “successful” censorship but also
eignty, and the pro-Russian minority, who praise and sup-
unintended outcomes that go against the censorship goals or
port the Russian Federation or Kremlin-backed separatists in
undermine the censors.
southeast Ukraine (Laruelle 2014, 2016). VKontakte is a pop-
ular source of entertainment in the post-Soviet space, and it
has a history of being an important source of pirated music The accessibility view on censorship
(Kiriya and Sherstoboeva 2015; Popkova 2019). Following the The accessibility approach emphasizes the accessibility of the
Kremlin’s turn toward strict internet regulations, VKontakte forbidden media and the apolitical evaluation of whether ac-
has fallen under great pressure and control from Russian au- cessing content is difficult. Instead of banning undesired content
thorities (Economic Times 2014; Pan 2017). While most of the entirely, the government may effectively limit user engagement
content on the platform is unrelated to politics, researchers, with the content by creating what Roberts (2018) refers to as
journalists, and authorities have argued that VKontakte is also friction: slowing down connections or blocking content in ways
a platform for pro-Russian propaganda (Dek et al. 2019; Van that can still be accessed through circumvention tools, such as
der Vet 2019; Volchek and Sindelar 2015), disinformation, VPN. By putting up minor barriers to censored content, the gov-
Russian surveillance, cyberattacks (Andrusieczko 2017), and ernment increases the incentives to use more easily available,
642 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

noncensored alternatives, while simultaneously avoiding the between forbidden and allowed information unclear, it is also
full-scale persecution of the masses and the potential political constantly changing. This leads to politically engaged indi-
backlash that may follow (Dickson 2016; Roberts 2018). viduals creating their own bottom-up explanations for which
Numerous studies indicate that citizens in authoritarian content is forbidden (and why) and internalizing a practice of
states respond to censorship by accessing available content in- self-censorship (1240–41).
stead (Chen and Yang 2019; Stockmann 2013). Interestingly, This literature strand explains not only successful censor-
the literature based on the accessibility view emphasizes that ship but also instances when such interventions backfire by
the users’ impatience and apolitical use of social media may provoking a political response. The political signals inherent in
also backfire—especially if the censored platform cannot be censorship stand as a possible driving mechanism behind the
replaced by a similar and freely available alternative to fill in previously mentioned Streisand effect. Here, the apparent
the gap (Hobbs and Roberts 2018). For instance, Hobbs and censorship generates more traffic toward the forbidden con-
Roberts describe how the abrupt Chinese government ban on tent by making people (1) more aware of the forbidden con-
the relatively apolitical Instagram increased the online traffic tent, (2) more curious, and (3) outraged (Jansen and Martin
toward prohibited and more political platforms such as Twitter 2015). While the first two aspects of the Streisand effect are not
and Facebook. They argue that this pattern is driven by a political per se, the third mechanism generates a backfire effect
“gateway effect.” The concept refers to a mechanism whereby due to a political response against the censorship policy itself.
the motivation to access a newly banned platform opens up for Using a more theoretical approach, Shadmehr and Bernhardt
skills (i.e., how to use VPN) that give access to other websites (2015) argue that censorship may cause the population to
that have long been censored (623–24). evaluate the regime more negatively in some instances, because
It is important to stress that this perspective does not ex- the lack of information may promote a belief that “there might
plain the response to censorship by referring to the users’ have been bad news that was censored” (280). In their empirical
political attitudes toward censorship or the government. On study, based on news media content data and approval surveys,
the contrary, this strand of literature is in line with the so-called Gläßel and Paula (2020) find that the German Democratic Re-
cute cat theory, which emphasizes that the search for enter- public’s censorship of information about the emigration crisis in
tainment—not political attitudes—is the main driving force 1989 backfired by causing outrage among people who detected
behind the citizens’ consumption of online content (Hobbs state misinformation through access to Western television.
and Roberts 2018; Zuckerman 2014, 624).
There are few studies that test the accessibility view on HYPOTHESES
censorship in a democratic context, likely because of the In this article, I use a multidimensional approach to analyze
literature’s focus on authoritarian or hybrid regimes. The view, user affiliations with Russia and Ukraine. First, I examine the
however, is highly relevant in democratic or semidemocratic political aspect of the affiliations by distinguishing between
societies. This is the case, precisely because liberal states are users with pro-Russian and pro-Ukrainian attitudes. The for-
ideologically bound to prioritize “softer” tools—such as ad- mer refers to praise and support for Russia; the latter, to support
justments to the cost of going online (in terms of time and of the Ukrainian national state. Second, I capture the social
effort)—over large-scale persecutions and fear of reprisal. aspect by examining social ties to individuals living in Russia.
My hypotheses and initial theoretical expectations are
The political-signaling view on censorship grounded in a political-signaling view for several reasons.
Whereas the accessibility view emphasizes the more prac- Unlike the more covert or subtle friction described in the ac-
tical and apolitical aspects of censorship, the political-signaling cessibility literature (using China as a case), the Ukrainian ban
view shifts the focus toward the political signal behind cen- against VKontakte was officially announced by the censors and
sorship as well as the role of political attitudes in the public widely discussed by the public. The censorship policy itself was
response toward such policies. The political-signaling approach described as a political move against Russia and a matter of
sees state-driven censorship and propaganda as a means of de- national security. The government, various journalists, and
marcating forbidden and undesired conduct. Governments can civil society groups framed the use of VKontakte as a political
therefore use censorship as a political signal to socialize citizens choice, a lack of patriotism, or even a part of the Russian in-
into the desired behavior—through implicit political messages, formation war against Ukraine. Using VKontakte in this heavily
fear, or uncertainty (Huang 2015; Stern and Hassid 2012). politicized context can be interpreted in the Ukrainian public
For instance, Stern and Hassid (2012) reveal how the Chi- as indifference toward patriotic ideals in times of war, whereas
nese government uses an atmosphere of uncertainty to govern compliance with the ban can be a sign of political loyalty to-
journalists, editors, and lawyers. Not only is the exact line ward Ukraine as a nation-state.
 Volume 84 Number 2 April 2022 / 643

If one sees censorship as a way for the Ukrainian govern- in Russia may be heavily embedded in Russian society, either
ment to signal that using VKontakte is both unpatriotic and through friends and family or by having previously lived in
undesired by the state, one would expect pro-Ukrainian users Russia. The user likely has high potential exposure (Hjorth and
to become less active on the Russian platform out of political Adler-Nissen 2019) to pro-Kremlin news, political narratives,
support for Ukraine and resentment toward the Russian and worldviews that are widespread among contacts in Russia
“aggressor state” or due to a fear of being stigmatized as “unpa- (Toal and O’Loughlin 2018). One would therefore expect these
triotic” by their Ukrainian peers. Following this line of thought, users to be more pro-Russian and more likely to resist the anti-
I expect the ban to succeed overall, because the majority of the Russian ban, which may essentially jeopardize their online ties
Ukrainian population is relatively supportive toward Ukraine’s to peers in Russia. In line with this, I supplement the hypothesis
sovereignty, while the pro-Russian individuals remain a minority above with the following hypothesis:
(Arel 2018). I therefore begin with the following hypothesis:
H1b. Users with fewer social ties to individuals living
H1. The censorship reduces online activity among in Russia are more affected by the censorship than
VKontakte users in Mainland. users with many social links to Russia.

It is important to note, however, that the accessibility view According to a report published by NATO StratCom, the
would predict a similar outcome for different reasons: the ban overall number of wall posts in Ukraine did fall substantively
would reduce the online activity by forcing users to install a following the ban (Dek et al. 2019, 58). It is worth noting that
VPN, regardless of their political orientation, and therefore pro-Russian posts increased, which, according to the authors,
slow down the log-in procedure or the connection. However, indicates “users moving to the pro-Russian infosphere” (51).
there is an important difference between the two views when it This alone is in line with my theoretical expectations based on
comes to the role of political attitudes. According to the logic the political-signaling view above. However, the report does
in the political-signaling view, users with strong Russian af- not use an explicit natural experiment design, nor does it com-
filiations would be less affected by the ban for at least two pare the effect of the ban among users with strong and weak
reasons. First, they may be less affected by the social pressure affiliations to Russia. The report presents little information on
to reduce their activity on the Russian platform because they the sampling strategy used to select users or posts, and it does
are already stigmatized as “pro-Russian” in the context of not document the clustering algorithms used to infer the
the armed Russian-Ukrainian conflict. Second, the Russian- ideology of posts. While the report may offer useful and in-
affiliated users would be more politically outraged by the dicative results, it is difficult to draw causal conclusions on the
“anti-Russian” policy and therefore more likely to resist a effect of censorship among different societal groups because of
ban that is heavily embedded in anti-Kremlin sentiment. This the above-mentioned issues. Further systematic inquiry is
theoretical expectation contrasts to the more pragmatic ac- needed if one is to understand how the effect is mitigated by
cessibility view on censorship, which would predict an equal individual-level factors.
decline in both groups due to increased costs of going online,
regardless of the users’ political attitudes toward Russia and CENSORSHIP AS AN EXOGENOUS SHOCK
Ukraine. Using this view as a point of departure, I formulate Before proceeding further, I first present arguments for why
the next hypothesis: the ban can be seen as an exogenous shock to the online media
ecology (i.e., the timing of the ban being largely unrelated to
H1a. Pro-Ukrainian users are more affected by the online activity on VKontakte). The executive order to ban a
censorship than pro-Russian users. list of Russian websites, including VKontakte, was signed by
Ukraine’s president at that time Petro Poroshenko on May 15,
As mentioned earlier, the Ukrainian government sees 2017, and publicly disclosed the next day, with no prior an-
VKontakte as not only platform for Russian surveillance but nouncement. Before the ban, there had been only a few waves
also a pathway of pro-Kremlin propaganda. From a theoretical of interest in the topic in February, when a Ukrainian public
point of view, VKontakte can be seen as a platform for “par- official suggested that the government should ban VKontakte,
ticipatory propaganda” in which the Ukraine-Russia conflict is with no indication of whether this would occur or when (see
socialized by ordinary users (Asmolov 2019) who actively take fig. A1; figs. A1–A3, F1, H1, I1, J1, K1, M1, N1, N2, Q1, and S1
part in producing and sharing propaganda with their friends are available online).
in Ukraine and abroad. The cross-national ties are key in this In this sense, the sudden ban can be seen as an exogenous
context. A Ukrainian user with many social ties to individuals shock to the online environment, with the exact timing being
644 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

largely unanticipated by the public and therefore “near ran- pany’s own application programming interface (API).1 All of
dom” from the perspective of the user. This is important, be- the user-level data were collected during August 2018, while
cause knowledge of the ban date would give users the oppor- the collection of wall posts took place in August–September
tunity to prepare by decreasing their activity in advance (e.g., of the same year (see app. T for an overview of the collected
migrating to other platforms) or increasing activity before the data). The data collection consists of six steps.
ban in defiance against the policy. This could bias the esti- First, I identify all of the cities in the VKontakte database
mated effect size of the actual ban. Because the exact timing of that are located in administrative regions (“oblast”) adjacent to
the actual censorship in 2017 is unexpected by the users, the the border separating Crimea and Mainland. In the second
period immediately before the ban may approximate the step, I use Google Maps API to automatically locate coordi-
counterfactual scenario in which the “treatment,” in the form nates for each city. I test the reliability of the automated ap-
of a ban, did not occur. From a research perspective, this proach by manually examining the results for each city. Of the
enables a natural experiment setting with a relatively clear 370 cities, only six have been misplaced in the automated
delimitation between the period before (control) and after process, which I have then corrected manually.
(treatment) the ban separated by a narrow window of up to Third, in order to filter out cities located more than
2 weeks for the policy to be fully enacted. 50 kilometers from the border, I use a geographic information
According to journalists, some users began reporting the systems approach to compute the distance between each city
effects of the ban as early as the evening of May 17 (UNIAN and the nearest point at the administrative land border sepa-
2017), although the ban was not fully implemented among all rating Crimea from Mainland. This relatively narrow delimi-
internet service providers at this stage. Because of technical tation is intended to maximize the regional similarity between
constraints and privacy concerns, it is not possible for me to the towns exposed to censorship in the treatment region and
pinpoint the exact hour of the ban for each individual user. For towns in the control region that have not been exposed, while
pragmatic reasons, I treat 00:01 a.m. on May 18 as the be- simultaneously keeping the distance bandwidth broad enough
ginning of the full-scale implementation. As I show in the to capture a sufficient number of public profiles from different
analysis, posting activity fell drastically on this day. political groups.
The Ukrainian government cannot implement censorship in In the fourth step, I sample the individuals by using the
all of its regions, because parts of its territory has been annexed VKontakte API to return data on up to 1,000 users from each
by Russia in 2014. I use this spatial variation to distinguish city within 50 kilometers of the border. I refer to this initial
between users who have not been exposed to the Ukrainian group of users as the “seed.”
ban, as they are located in Russian-occupied Crimea, and those In the fifth step, I develop a crawler to initiate a snowball
who were likely exposed to the ban, as they are located in sampling with the seed group as the initial starting point. In
Mainland (i.e., on the nonoccupied side of the Crimean border this crawl, I first download the metadata for the friends of the
in Kherson Oblast). seed group, only keeping those that live in cities within the
I use Crimea for comparison—and not neighboring regions 50 kilometer distance from the border according to the user’s self-
countries like Russia, Belarus, or Poland—because of its rela- registered data. I then download the metadata for the friends’
tive similarity to adjacent Ukrainian regions as well as a clear friends, continuing the same procedure six times. Because of
border separating censored and noncensored territories, un- the highly interconnected nature of both online and offline
like the frontline in the war-torn Donbass region (see app. G; networks, this snowballing strategy yields a nearly complete
apps. A–T are available online). It is important to note that the sample of 109,191 users within 50 kilometers of the border (as
geographical border separating the censored and noncensored well as their 3.2 million friends from around the world with
users is not random. For this reason, the spatial delimitation self-registered locations). Approximately 65,213 of the users
does not in itself enable a natural experiment research design, within the 50 kilometer border logged on to the website for the
unlike the as-if-random, temporal “line” separating the censored last time at least 30 days before the ban or during the period
and noncensored time period. However, the comparison of on- after the ban. Like the rest of the data, the last log-in dates for
line activity across geographic space adds both valuable nuance the public accounts are collected using VKontakte’s API. This
and validation to the comparison of online activity over time. information is also a highly visible feature of the public profile

DATA 1. For this purpose, I have used and modified code from the vkR package
To compare the change in VKontakte before and after the ban, in R developed by Dmitriy Sorokin: https://cran.r-project.org/web/packages
I collect publicly available data on VKontakte from the com- /vkR/index.html.
 Volume 84 Number 2 April 2022 / 645

and easily accessible by all VKontakte users. In order to ensure post on their own walls before the May 18, 2017, implemen-
that the data are not biased toward users with high posting tation date. Descriptive statistics for the final subsample used
activity after the ban, this sampling step is carried out inde- in the analysis are available in appendix A. The findings re-
pendently of the users’ posting behavior (the data therefore main robust when including users without a single public post
also include accounts with no public posts). The distribution of before the ban (see app. Q). In total, the subsample in the analy-
these users on both sides of the border between Mainland and sis section includes 1,067 pro-Russians, 1,112 pro-Ukrainians,
Crimea is illustrated in figure 1. and 845 users from both sides of the border. While the random
In the final step, I select a smaller subsample of 3,553 active users are not necessarily politically neutral, they are likely less
users and all 2,127,398 public posts from their respective walls pro-Russian or pro-Ukrainian than those who publicly expose
to capture individual-level online activity before and after the their political affiliation. I use the information about the users last
ban. I select the individuals using a stratified sampling, to log-in date to examine whether users accessed their VKontakte
ensure that I have users in the different political groups from account after the ban.
both sides of the border (approximately 590 users per group in The data from VKontakte are used to operationalize user
Mainland and 660 in Crimea) as well as randomly selected activity, geographic location, social ties to Russia, and political
users from the pool of 65,213 active profiles. I then remove attitudes toward Russia and Ukraine. These variables are de-
51 profiles that have posted on average more than three posts scribed a follows.
daily (26,456 posts in total) in the 90 days before the ban User Activity is operationalized as the number of public
(treatment), in order to filter out extreme outlier accounts that posts written by the users on their own profile walls. In this
are likely automated (i.e., inauthentic bot accounts; Varol et al. sense, user activity is different from mere access to an account.
2017). These profiles account for less than 1.5% of the original Users may, for instance, circumvent the ban through VPN
subsample. The findings in this article remain robust when tools to passively access their accounts without generating
also including the hyperactive users. While VKontakte API any public posts on their walls. Geographic Location is oper-
provides data on an account’s last log-in time, the website does ationalized using the profiles’ self-reported information on
not disclose when the account was created. In order to ensure “current city” in August 2018. Social Ties to Russia is measured
that the profiles existed before the ban, I further delimit the as the proportion of users’ friends who have publicly indi-
data to the 3,024 users who have uploaded at least one public cated that they live in Russia.

Figure 1. VKontakte users within 50 kilometers of the Crimean border. All of the users have logged in within 30 days of the ban. Color version available as an
online enhancement.
646 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

Political Attitude toward Russia and Ukraine is opera- fect of the ban when using Crimea as a control. In the other
tionalized as follows: pro-Russian users as those who only words, the term reflects the difference between the change in
follow pro-Russian VKontakte community pages, while pro- the mean number of posts per day in Mainland (treatment
Ukrainian users are those who only follow pro-Ukrainian com- group) and Crimea (control group). The main hypothesis
munities (see apps. B–D for details on sampling and coding predicts that the ban has a negative effect on posting activity
of the communities). While the military conflict is often dis- (b3 Banit #Mainlandit ! 0). Additional control variables in-
cussed in Russia in terms of an ethnic divide (Shklovski and clude total number of VKontakte friends per August 2018 and
Wulf 2018), speaking Russian does note equate with Russian a dichotomous variable indicating whether the post was written
identity, and Russian identity among Ukrainian citizens does during the weekend as well as the users’ self-reported gender.
not necessarily imply pro-Russian attitudes in the current To test whether the effect is mitigated by political affilia-
Russia-Ukraine conflict (Arel 2018). tions and social ties to Russia, I create two additional DD
models only for pro-Russian and pro-Ukrainian users from
DIFFERENCE-IN-DIFFERENCES SPECIFICATION Mainland. Here, I replace Mainland with (1) a binary variable
In this section, I describe the empirical setup for measuring the indicating whether the user has an above-median or up-to-
effect of the ban on online activity. All of the models in this median proportion of VKontakte friends and (2) political
article are based on individual-level data, where each user-time affiliation toward Russia and Ukraine, respectively. If the ef-
observation represents a number of wall posts uploaded by the fect of the ban is mitigated by these factors, the respective
individual users on a given day. interaction terms will be significantly different from zero.
I use a difference-in-differences (DD) approach (Card and In the final step, I examine the difference between the effect
Krueger 1993) to estimate the average effect of the ban on user size of the ban among the various groups, by introducing
activity among the individuals in the subsample living in Main- Crimea as a control. The literature refers to this approach as
land within 50 kilometers of the border. Minor fluctuations in the difference-in-difference-in-differences, or triple differences
posting activity are common on both sides of the border, both (DDD), method (Yelowitz 1995). More specifically, I expand
before and after the ban (see fig. A2). One could reasonably the equation above by introducing a three-way interaction
assume that online activity in Mainland would have experi- between Ban, Mainland, and a dichotomous variable (Pro-
enced a small and gradual decline—even if the ban would not Russian) indicating whether or not the user is pro-Russian:
have occurred. This means that a simple comparison of posting
activity before and after the ban only in areas exposed to cen- yit p a 1 b1 Banit 1 b2 Mainlandit 1 b3 Pro‐Russianit
sorship (treatment) runs the risk of overestimating the effect 1 b4 Banit # Mainlandit 1 b5 Banit # Pro‐Russianit
size. For these reasons, I introduce users from noncensored
1 b6 Banit # Mainlandit # Pro‐Russianit 1 dX 1 εit :
Crimea as an additional control.
ð2Þ
The advantage of using DD design in this context is that
it allows treatment and control groups to be different, since This triple difference is indicated by the interaction term
the method is not based on the assumption that the treatment Banit # Mainlandit # Pro‐Russianit , which can be broken down
(the ban) is randomly assigned. The key argument for using into several parts. The interaction Banit # Pro‐Russianit shows
this design is the similarity in time trends between the two the difference between the change in mean posting activity
groups—that the Crimean trend approximates the counter- among pro-Russian and pro-Ukrainian profiles in Crimea
factual scenario in Mainland where the ban did not occur. (which is in the reference category). This provides an estimate
I specify the DD regression using the following ordinary of how much more pro-Russian users are affected by the ban
least squares model: than the pro-Ukrainian ones. By introducing Mainlandit to the
interaction term, the estimate shows the difference between the
yit p a 1 b1 Banit 1 b2 Mainlandit 1 b3 Banit
ð1Þ estimate for Banit # Pro‐Russianit in Mainland (treatment) and
# Mainlandit 1 dX 1 εit : Crimea (control). If, for instance, the difference between the
In this equation, yit indicates the mean number of posts up- pro-Russian and pro-Ukrainian users is identical on both sides
loaded by user i during day t, where the earliest implementation of the border, the interaction term will be close to zero. I use
date (May 18, 2017) is standardized as 0. The binary variable the same approach by replacing Pro-Russianit with a binary
Banit indicates whether the posts are uploaded by the user before variable, Friends in Russia, which indicates whether the users
or after the ban, and Mainland reflects whether the user lives in have an above-median (Friends in Russia p 1) or up-to-
Mainland (Mainlandit p1) or Crimea (Mainlandit p 0). The median proportion of VKontakte friends in Russia (Friends
interaction term, b3 Banit #Mainlandit indicates the main ef- in Russia p 0). Because users in Crimea have on average a
 Volume 84 Number 2 April 2022 / 647

higher proportion of VKontakte friends in Russia, I compute mean effect of the ban on posting activity in Mainland. In this
two separate median values for users in Mainland and Crimea, part of the analysis, I use a subsample of 3,024 profiles from
respectively.2 I use cluster robust standard errors for all of pro-Russian, pro-Ukrainian, and random users from both
the DD and DDD models, in order to take into account the sides of the border. Each of the 547,344 user-day observations
structure of the panel data, where there are multiple user-time reflects the number of posts uploaded by the respective in-
observations for the same individuals across time. dividual on a given day within the bandwidth of 90 days be-
fore and after the cutoff date (181 days in total), standardized
as 0. Figure 2B shows the observed mean number of posts per
LAST LOG-INS
user for all of the 3,024 users in the subsample. The figure
Before proceeding with the analysis of daily posting activity,
illustrates the drastic reduction in the daily posting activity for
I first examine whether users log in to their accounts after
all Mainland users. In comparison, the ban is accompanied by
the ban or whether they last did so during the ban period. “Last
a relatively small, continuous decline in Crimea.
log-in,” in this case, refers to the last time a nondeleted account
The effect size of the sudden censorship policy in the DD
logged in as of August 2018, the data collection period. Of the
model is reflected by the interaction term Ban # Mainland in
109,191 accounts from both sides of the border in this study,
figure 3 and table I1 (tables A1, E1, E2, I1–I3, K1, K2, L1, M1,
65,213 (59.7%) last logged in during the 30 days before the ban
M2, O1, Q1, Q2, R1, and R2 are available online). The ban
and the data collection date. The remaining 40.3% logged in
reduced the posting activity for the average user by 0.8 posts
for the last time 1 month before the ban or earlier.
per week or 0.114 posts per day (t p 212:4075, p ! :01) in
The findings suggest that the vast majority of the non-
Mainland, when using a 30 day bandwidth before and after
deleted accounts continued to log back in after the ban. Ap-
the ban (61 days in total). In other words, Mainland online
proximately 94.1% of the 23,506 active users in Mainland
activity (treatment group) declined by an additional 0.114 daily
logged in at least once after the implementation date, and
mean number of posts for the average user, compared to the
90.8% continued logging in more than 30 days after the ban.
change in Crimea (control group).
For users in Crimea, the figures are 99.0% and 98.3%, re-
This effect size is substantively large. The ban decreased the
spectively (see app. H for an in-depth comparison between
daily mean number of posts per user by 45.42%, according to
Mainland and Crimea). Of all the 1,433 users in the subsam-
the most conservative estimate (based on 30 day bandwidth).
ple from Mainland who have posted at least once before the
When using the 90 day bandwidth, the decline is equivalent
ban, 93.4% continued logging in more than 30 days after the
to 47.01% of the preban daily mean number of posts.
ban, even though the policy remains both legally and tech-
nically in place to this day. For pro-Russian, pro-Ukrainian,
THE DIFFERENCE IN EFFECT SIZE
and random users in Mainland the proportions are 96.4%,
The findings above confirm the first overall hypothesis: the ban
91.8%, and 91.4%, respectively. The pattern is illustrated in
reduces the online activity on VKontakte in Mainland, also
figures 2A and A3. The difference between the proportions for
when examining posting activity. This leads to the following
the pro-Russian and pro-Ukrainian users is statistically sig-
question: To what extent is the censorship effect mitigated by
nificant (p ! :05) albeit not substantively large. This indicates
different factors?
that a large percentage of users have found a way of avoiding
Figure 4 gives a descriptive overview of the observed mean
the censorship, some through freely available VPN services.
number of posts per user for each day for the different groups
However, these numbers do not include the change in posting
and all of the users combined. The change in posting activity
activity among the users who have bypassed censorship. I
appears strikingly similar when visually comparing user groups
therefore turn to an analysis of a smaller subsample of users
with different (political and social) affiliations to Russia and
from both sides of the border in order to estimate the impact
Ukraine. In this section, I use DD and DDD models to test
of the ban on their daily posting activity.
whether the difference is statistically significant and substan-
tively large when including control variables.
THE OVERALL EFFECT OF THE BAN I turn to this question by testing hypothesis 1a, which
How did the Mainland posting activity change compared to predicts that the anti-Kremlin ban affects pro-Ukrainian users
northern Crimea within just 50 kilometers, where there was more than pro-Russian users. This part of the analysis is limited
no ban? To answer this question, I use DD to estimate the to pro-Russian and pro-Ukrainian users (without random
profiles), in order to enable the comparison between the two
2. The findings remain robust when using a single median value for groups. The variables of interest for the DD and DDD models
profiles on both sides of the border. are illustrated in figure 5.
648 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

Figure 2. Last log-ins and posting activity. A, Proportion of users who have access to their VKontakte profile. The users are considered to have access to VKontakte
until their very last log-in date. All of the 3,024 accounts are assumed to have existed throughout the entire period before the censorship. This can be confirmed for at
least 98.84% of the accounts (see the description in fig. A3 for details). Note that the decline begins before the ban. This is because the data are based on last log-
ins; that is, some users log in for the last time, not knowing that they will be locked out by the sudden ban. B, Posting activity for all 3,024 accounts. Horizontal lines
reflect observed means for the respective periods and groups. Color version available as an online enhancement.

I find no statistically significant difference between the two Ukrainian users is reflected by the interaction term Ban#
groups when using DD and a significance level of 5% to ex- Pro-Russian#Mainland in figure 5 and table I2 (model 5).
amine the data only from Mainland. The difference for the According to the most conservative estimates—based on a
DD models is indicated by the coefficient for the interaction 90 day bandwidth—the effect size of the ban is 0.004 posts
term Ban#Pro-Russian in the output in figure 5 and table I2 greater among pro-Russian users than pro-Ukrainian ones
(model 3). While the difference is statistically significant at (t p 20:3202, p p :7488), contrary to the hypothesis. While
a 10% level for the 90 day bandwidth, these estimates indi- this difference appears even greater, when narrowing the
cate that pro-Russian users are more affected by the ban than bandwidth down to 30 days, it remains statistically insignifi-
pro-Ukrainian users. Here, the ban reduced the average daily cant. The data therefore provide little evidence in support of
activity among pro-Russian users by 0.017 posts more than hypothesis 1a.
their pro-Ukrainian counterparts (t p 21:8229, p ! :1), con- I find similar results when testing hypothesis 1b, which
trary to the hypothesis. predicts that individuals with fewer social ties to users in
Similarly, the results show little evidence supporting the Russia are more affected by the ban than those with extensive
hypothesis when adding Crimea as a control to the DDD social ties linking to the country. I examine the effect by com-
model. Here, the difference between pro-Russian and pro- paring users with an above-median percentage of VKontakte

Figure 3. Change in mean number of posts per day after the ban, 95% confidence intervals. Reflects models 1 and 2 in table I1, which are based on data from
all 3,024 users (including pro-Russian, pro-Ukrainian, and random users).
 Volume 84 Number 2 April 2022 / 649

Figure 4. Change in mean number of posts per day after the ban, for all users and the split sample. Posting activity for all users, as well as the split samples:
pro-Russian, pro-Ukrainian users, as well as those with up-to-median and above-median proportions of friends in Russia. Horizontal lines reflect overall
means for the respective periods and user groups. Color version available as an online enhancement.

friends living in Russian to those with an up-to-median pro- as an additional control in the DDD model, by including the
portion. There is no statistically significant difference in how following interaction term: Ban#Friends in Russia # Main-
much the two groups are affected when only using data from land. According to the most conservative DDD estimate, the
Mainland. This is reflected by the Ban # Friends in Russia ban has reduced online activity by 0.0185 posts (t p 20:8212,
interaction term in figure 5, when using a bandwidth of 30 p p :4115) more among users with above-median proportion
and 90 days. The pattern is consistent when introducing Crimea than those with fewer friends in Russia—contrary to the

Figure 5. Change in mean number of posts per day after the ban, with 95% confidence intervals. Reflects models in tables I2 and I3, which are based on data
from users with political affiliations (not random users). The order of the interaction terms is based on the model numbers, so that the first interaction term
(top) is based on model 1 and the fifth interaction term (bottom) is based on model 5.
650 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

theoretical expectation. However, this difference remains sta- users and those with relatively many social ties to Russia are
tistically insignificant. These results provide little evidence more likely to comply with the anti-Russian ban by reducing
supporting hypothesis 1b. their online activity on VKontakte.
I find similar results when rerunning the models behind The potential difference between the two groups cannot be
figure 5 with user-level fixed effects (see fig. I1) or negative explained by fears of legal reprisals or social stigma alone.
binomial regressions instead of ordinary least squares. Fur- More than 90% of individuals in the subsample continued to
thermore, the findings remain robust when testing for long- log in to their public accounts at least one month after the ban
term effects by expanding the time period, using regression without hiding their last log-in date. This means that their
discontinuity in time models instead of DD, using different friends, strangers, and authorities alike can see that the re-
independent variables (including strength of ties to Russian spective users continued using VKontakte at least one month
users), and reiterating the analysis on users from Kyiv. The after the ban. This information is visible even if the users do
results based on DD and DDD remain the same when delim- not write any posts.
iting the analysis to users who managed to technically cir-
cumvent the censorship by logging in at least 30 days after the
ban. In other words, the censorship reduced VKontakte activity IMPLICATIONS
even among users who were capable of circumventing the ban. I argue that the behavioral response to censorship in Ukraine
I validate the political affiliation variable by examining whether is largely in line with the competing accessibility view on cen-
users labeled “Pro-Russian” are more connected to Russia than sorship, represented by scholars such as Hobbs and Roberts
“Pro-Ukrainian” users through friendship networks, reposts, (2018). Seen from this theoretical perspective, the response to
and use of Russian websites, as one would expect. I find that this censorship is driven by consumer impatience and the in-
is the case. I conducted a placebo test by reiterating the analysis creasing costs of going online and to a lesser extent political
on the same dates one year before the ban. By doing so, I show affiliations with Russia or Ukraine. This view explains why
that the results are not driven by reoccurring seasonal events. pro-Russian users and those with potential exposure to Rus-
Appendix J provides an in-depth overview of the robustness sian information networks are no more likely to resist the
and validity tests. anti-Russian ban than pro-Ukrainian users or those with few
social ties to Russia.
The Ukrainian ban offers a hard case for the practical ac-
A PRAGMATIC RESPONSE TO CENSORSHIP cessibility view on censorship. Not only does the government
The vast majority of pro-Russian and pro-Ukrainian users actively politicize Russian media as a means of information
continue logging onto the forbidden website despite the ban. warfare against Ukraine, the users are also encouraged to
Pro-Russian users are slightly more likely to maintain passive demonstrate loyalty to their country by ceasing to use services
access to the platform than their pro-Ukrainian counterparts. from the aggressor state. Yet the findings suggest that the
However, the difference is not substantively large, and they are mechanisms described by Hobbs and Roberts (2018) and
at least as affected when it comes to online activity. Why do Roberts (2018) in their study of China offer a powerful ex-
users, who can legally access VKontakte and have the tech- planation even in a heavily politicized context of war, where
nical know-how to do so, still choose to shift their attention the forbidden media itself is portrayed as a weapon.
away from the Russian platform? Whereas Hobbs and Roberts (2018) discuss the incentive
From the political-signaling point of view, the decision to to acquire the skills to access forbidden websites via VPN, the
either contest or comply with the government ban is driven by Ukrainian case shows that the mechanisms behind the ac-
political attitudes toward the regime. From this perspective, cessibility view may hold even when the majority of the po-
the ban is effective because it actively signals that using the litically interested users find a legal way to circumvent censor-
online services provided by the “aggressor state” is unpatriotic ship free of government persecution. The practically oriented
and a potential threat to national security. In other words, the and pragmatic user has an incentive to log on to VKontakte
underlying reason for reducing activity on the platform is occasionally, either to respond to unanswered messages or to
more political than practical. One would therefore expect pro- check for new ones. However, free VPN services often add
Ukrainian individuals and those with fewer social ties to friction by slowing down the connection or providing secure
Russia to be more compliant with the anti-Kremlin ban on browsing only for a limited traffic volume. For example, at
Russian social media by reducing their online activity. The least three out of the five VPN services that the VKontakte
findings present little evidence for this view; on the contrary, team has publicly recommended for circumventing the Ukrai-
if there is a difference between the two groups, pro-Russian nian ban (VKontakte 2017) require individuals to pay a fee for
 Volume 84 Number 2 April 2022 / 651

a fast connection or unlimited browsing. One of the free options still be highly vulnerable to government censorship. This is
requires individuals to use the Opera browser. This introduces important, because it emphasizes that the effects of censor-
additional obstacles for those who are used to Google Chrome, ship cannot be evaluated by examining the technical ability of
Firefox, or other widespread browsers. Even in this case, some the population to bypass censorship alone. The findings also
bloggers have complained that their connection began to lag serve as a reminder to political scientists that practical and
more when using the Opera VPN—right after users from relatively pragmatic considerations may play a more impor-
Ukraine began to switch to the service (help-wifi.com 2018). tant role in terms of how the users respond to online cen-
The cost of logging in increases dramatically, if the user sorship than their political affiliations toward the censoring
chooses to move away from the free options in order to ac- state or to its foreign “enemy”—even if the opposing political
quire a faster and more efficient service. A paid VPN service may groups are involved in war against each other.
cost an additional US$3–US$7 per month—a sizable amount How generalizable are these findings? As mentioned pre-
in Ukraine. However, VPN clients may potentially complicate viously, the results are in line with the accessibility view on
the log-in procedure even if they succeed to provide a fast information restrictions, which has been used to describe on-
connection. A pro-Ukrainian resident in Kyiv explained to me line censorship in China. The findings in this article suggest the
in an interview why he no longer uses VKontakte on a daily pattern can be generalized to a more democratic and politi-
basis: the VPN on his PC required more clicks and slowed cized context, such as the war in Ukraine. Following the logic
down the connection by a few seconds. Interestingly, the same of this accessibility view, one could expect a similar pattern in
person could enter VKontakte automatically on his smart- other countries if at least three conditions are met.
phone without extra clicks, likely due to a preinstalled VPN. First, the ban must be technically implemented in a way
Despite this, he still decided to stop using VKontakte as his that significantly complicates and slows down access to the
main platform because he felt it was too troublesome to manage forbidden source. This is not always the case. For example, the
his account only from a smartphone and not being able to do so Russian authorities arguably failed to do so in their ban on
unhindered through a PC. This evidence is anecdotal and calls the Telegram app in 2018 (see app. P). Because Telegram
for more (qualitative and quantitative) research on the user moved its service to Google and Amazon domains (i.e.,
experience of banned platforms. Nevertheless, the story em- through “domain fronting”), users could continue accessing
phasizes the potential complexities of managing forbidden ac- the website with relative ease during prolonged periods. The
counts across multiple devices and operating systems. service was officially unblocked in 2020.
Because of the friction created by the censorship policy, the Second, users must have unhindered access to similar ser-
impatient user has an incentive to switch to a cheaper alter- vices that can partly replace the banned product. In the case
native that is unaffected by the ban: Facebook. The American of Ukraine, VKontakte can be replaced with highly similar
website has many of the same functions as its Russian com- alternative: Facebook. Conversely, if Russian authorities were
petitor but can be accessed without VPN. While the data in to ban Facebook, as they have previously threatened to do
this article do not reveal the extent to which the users in the (Doffman 2019), the service could be replaced with VKontakte.
sample have migrated to other social media platforms, aggre- In the example above, Telegram was likely difficult to replace
gated data from online traffic monitors suggest that this may in Russia with a more “cooperative” platform, because its core
have been the case. According to data from SimilarWeb (an selling point at the time was precisely its unwillingness to share
analytics firm), used in NATO StratCom‘s report, VKontakte’s user data with the authorities. If there is no popular replace-
ranking fell from the first most visited website in Ukraine on ment for the banned platform the impatient users, some of whom
March 18, 2017, to the fifth most visited site on August 14, 2018 are addicted to the product, would have had a much higher
(Dek et al. 2019, 42). Facebook, however, jumped from the incentive to circumvent the ban (Hobbs and Roberts 2018).
eighth to fourth place in the same period. This can be viewed Finally, I theorize that similar bans may be more effective,
as a great success, seen from the perspective of the Ukrainian if the forbidden platform is already widely known. As pointed
government. As mentioned earlier, the purpose behind the ban out by the literature on the Streisand effect, a ban may backfire
is not necessarily to reduce social media behavior as a whole by making the general public more curious and aware of the
but to push users toward other media alternatives—away from forbidden content. This was less likely during the Ukrainian
a social media platform that provides Russian authorities with ban in 2017, precisely because VKontakte, unlike Telegram in
surveillance opportunities while showing no commitment to Russia during 2018, was already a mundane household name
curb pro-Kremlin disinformation campaigns. and a part of everyday life for the majority of online users.
Because of these factors, even users with access to the While these theoretical perspectives may offer a guideline for
“freedom technology” needed to circumvent online bans may future inquiries, more cross-national research is needed to
652 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

empirically examine when and how censorship succeeds in how the users have responded to the ban, it is still possible that
one context and not the other. these factors influence the consumption of Russian services in
the long run, gradually and independent of sudden shocks to
LIMITATIONS the media ecology.
For ethical reasons, I measure online activity by using only
publicly available posts from public accounts and not private CONCLUSION
posts or messages. This article does not argue that politics do This article shows that online censorship may work even
not matter in the citizens’ response to censorship or that po- when users have the technical and legal means to circumvent
litically engaged users are less affected by the ban than those it. The Ukrainian government succeeded in reducing the over-
who are disinterested in politics. Instead, the study is focused all online activity on VKontakte, which it perceives as a plat-
on the difference between users with strong and weak (political form for foreign propaganda and a tool for surveillance used
and social) affiliations with Russia. This question is pivotal to against Ukraine in times of war. Even a minor increase in the
understanding the extent to which the ban has backfired by costs of entering on the banned website is enough to signifi-
affecting one affiliation group more than the other. cantly shrink online activity, as long as there are more ac-
The findings suggest that users with either pro-Russian or cessible alternatives.
pro-Ukrainian affiliations are on average less affected by the The data support the accessibility view on censorship,
ban than random users from the same region. One possible which argues that the users respond to censorship primarily
explanation is that users with political affiliations are on av- on the basis of the costs of going online (in terms of time
erage more active online (see app. A) and therefore more and effort), not their political attitudes toward the regime.
invested in the platform. More research is needed to validate The anti-Russian ban, meant as a response to Ukraine’s con-
this view. Although Ukraine’s measure against the “unpatri- flict with Russia, affected pro-Russian users as much as pro-
otic” Russian platform may have made political users on both Ukrainian ones in terms of online activity. Similarly, the ban
sides more prevalent in relative terms, it affected the pro- is at least as likely to reduce the activity among users with
Russian users at least as much as the pro-Ukrainian ones. It is high levels of potential exposure to pro-Russian content, mea-
important to note that the pro-Russian users in the data set sured as social ties to Russia, as those with low levels. In other
publicly follow pro-Russian communities in a time of war and words, the ban did not make VKontakte more pro-Russian in
are therefore likely to be relatively extreme in their political relative terms.
attitudes. More research is needed to test whether the findings The results are favorable from the perspective of the cen-
hold among individuals with only slightly pro-Russian views sor, who wishes to combat foreign propaganda and disinfor-
or individuals who do not have any visible affiliation with mation by using one of the most drastic countermeasures
Russia in the online realm. Similarly, this research does not available. However, this article paints a more concerning pic-
exclude that the effects of the ban may vary depending on ture for those who see censorship as a threat to democracy.
whether the users access VKontakte through a static connec- This study contributes to the growing evidence suggesting
tion or a mobile connection on a smartphone. that states can greatly regulate online consumption by setting
Furthermore, the study does not shed light on whether up minor obstacles between the impatient user and the banned
the ban has affected the consumption of pro-Kremlin con- service. The Ukrainian case indicates that this may be true even
tent among Ukrainian users offline or whether it changed in a relatively democratic context and among users who al-
the popularity of different political topics (see Dek et al. 2019). ready have the technical means to bypass censorship and can
In line with this, the findings do not reveal the extent to which the do so free of persecution.
ban has affected political attitudes among ordinary citizens. Given these findings, one would expect censorship in other
Instead, the results show that the Ukrainian authorities countries to lead to a similar outcome. A sudden, large-scale
succeeded in achieving an important goal in their antipro- ban would not make the opposition more predominant on the
paganda policy: to reduce online activity on Russian social forbidden platform compared to the proregime movement.
media. From the perspective of the Ukrainian government, Similarly, if Russia were to use its newly upgraded censorship
this is an important achievement because it sees Russian social infrastructure to ban Facebook to prevent foreign influence,
media as a platform for foreign propaganda, surveillance, one would expect the ban to be successful from the point of
disinformation, and a national security threat. view of the government if Russians were to respond in a similar
The findings in this study are limited to instances when manner as Ukrainians have. Understanding the mechanisms
users respond to sudden censorship on social media. Although behind such interventions is important if one is to shed light
political and social affiliations have not played a major role in on how states police cyberspace. This calls for more research
 Volume 84 Number 2 April 2022 / 653

on the individual-level factors that may mitigate the effects of Doffman, Zack. 2019. “Facebook, Instagram and YouTube Will Now Be
Banned, Russians Warned. Forbes, September 13. https://www.forbes.com
censorship, together with cross-national comparative studies
/sites/zakdoffman/2019/09/13/putin-now-plans-100-facebook-instagram
of how and when large-scale social media shutdowns are -and-youtube-bans-russians-warned/ (accessed October 26, 2020).
met with political resistance among citizens rather than user Economic Times. 2014. “Pavel Durov Left Russia after Being Pushed Out.”
compliance. Economic Times, April 22.
Edwards, Lilian. 2009. “Pornography, Censorship and the Internet.” In Lilian
Edwards and Charlotte Waelde, eds., Law and the Internet. Portland, OR:
ACKNOWLEDGMENTS Hart, 623–70.
I would like to thank the JOP editors and anonymous re- Esberg, Jane. 2020. “Censorship as Reward: Evidence from Pop Culture
Censorship in Chile.” American Political Science Review 114 (3): 821–36.
viewers for their detailed comments and invaluable work. I am Fiorentino, Michael-Ross. 2018. “France Passes Controversial ‘Fake News’ Law.”
very grateful to Rebecca Adler-Nissen, David Dreyer Lassen, Euronews, November 22. https://bit.ly/3bcPZsm (accessed October 26, 2020).
Sune Lehmann, Margaret E. Roberts, Sarah Oates, Anders Gläßel, Christian, and Katrin Paula. 2020. “Sometimes Less Is More: Cen-
Wivel, Mikhail Suslov, Asmus Leth Olsen, Valentina Shapo- sorship, News Falsification, and Disapproval in 1989 East Germany.”
American Journal of Political Science 64 (3): 682–98.
valova, Frederik Hjorth, Benjamin Carl Krag Egerod, David
help-wifi.com. 2018. “Pochemu internet s vklychennim vpn ‘tupit’ i ‘tormozit’?”
Szakonyi, Joshua A. Tucker, James Pamment, Alicia Fjällhed, https://help-wifi.com/reshenie-problem-i-oshibok/pochemu-internet-s
Lene Hansen, Kasper Møller Hansen, Barbara Maria Pio- -vklyuchennym-vpn-tupit-i-tormozit/ (accessed October 26, 2020).
trowska, Arjen van Dalen, Marie Heldgaard Seestedt, Jeppe Hjorth, Frederik, and Rebecca Adler-Nissen. 2019. “Ideological Asymmetry
in the Reach of Pro-Russian Digital Disinformation to United States
Vierø, Frederik Kristoffer Kjøller Larsen, Gregory Eady, and
Audiences.” Journal of Communication 69 (2): 168–92.
Simon Polichinel von der Maase for their thorough and helpful Hobbs, William R., and Margaret E. Roberts. 2018. “How Sudden Censorship
feedback. I would also like to thank everyone who discussed Can Increase Access to Information.” American Political Science Review
this article during events at the 2019 American Political Sci- 112 (3): 621–36.
ence Association Annual Meeting, Social Media and Political Howard, Philip N., Sheetal D. Agarwal, and Muzammil M. Hussain. 2011.
“When Do States Disconnect Their Digital Networks? Regime Responses to
Participation Lab, Copenhagen Center for Social Data Science, the Political Uses of Social Media.” Communication Review 14 (3): 216–32.
and the PhD seminar as well as the workshops at the Uni- Huang, Haifeng. 2015. “Propaganda as Signaling.” Comparative Politics 47 (4):
versity of Copenhagen. 419–44.
Interfaks. 2018. “Zhitelya Peterburga sobralis napravit na prinuditelnoye
lechenije za anekdot ‘Vkontakte.’” Interfaks, August 20. https://www.interfax
REFERENCES .ru/russia/625841 (accessed October 26, 2020).
Andrusieczko, Piotr. 2017. “Ukraine Blocks Access to Russian Internet.” Jansen, Sue Curry, and Brian Martin. 2015. “The Streisand Effect and Cen-
Outriders, May 18. https://bit.ly/2QGYW5j (accessed October 26, 2020). sorship Backfire.” International Journal of Communication 9:656–71.
Arel, Dominique. 2018. “How Ukraine Has Become More Ukrainian.” Post- Kalathil, Shanthi, and Taylor C Boas. 2010. Open Networks, Closed Regimes:
Soviet Affairs 34 (2–3): 186–89. The Impact of the Internet on Authoritarian Rule. Washington, DC:
Asmolov, Gregory. 2019. “The Effects of Participatory Propaganda: From Carnegie Endowment.
Socialization to Internalization of Conflicts.” Journal of Design and Sci- Kern, Holger Lutz, and Jens Hainmueller. 2009. “Opium for the Masses: How
ence (6). https://doi.org/10.21428/7808da6b.833c9940. Foreign Media Can Stabilize Authoritarian Regimes.” Political Analysis
BBC. 2017. “Twitter Bans RT and Sputnik Ads Amid Election Interference Fears.” 17 (4): 377–99.
BBC News, October 26. https://bbc.in/33DJ1dc (accessed October 26, 2020). Kiriya, Ilya, and Elena Sherstoboeva. 2015. “Piracy and Social Change: Rus-
Boichak, Olga, and Sam Jackson. 2020. “From National Identity to State sian Media Piracy in the Context of Censoring Practices.” International
Legitimacy: Mobilizing Digitally Networked Publics in Eastern Ukraine.” Journal of Communication 9:839–51.
Media, War and Conflict 13 (3): 258–79. Kiselyova, Maria, and Alessandra Prentice. 2017. “Russian Social Media Site
Bunn, Matthew. 2015. “Reimagining Repression: New Censorship Theory Tells Ukrainians How to Dodge Web Block.” Reuters, May 17. https://
and After.” History and Theory 54 (1): 25–44. reut.rs/2UsKW04 (accessed March 11, 2020).
Card, David, and Alan B. Krueger. 1993. “Minimum Wages and Employment: Lanoszka, Alexander. 2016. “Russian Hybrid Warfare and Extended Deter-
A Case Study of the Fast Food Industry in New Jersey and Pennsylvania.” rence in Eastern Europe.” International Affairs 92 (1): 175–95.
NBER Working paper 4509, National Bureau of Economic Research, Laruelle, Marlene. 2014. “Russian Nationalism and Ukraine.” Current History
Cambridge, MA. 113 (765): 272–77.
Chen, Yuyu, and David Y. Yang. 2019. “The Impact of Media Censorship: 1984 Laruelle, Marlene. 2016. “The Three Colors of Novorossiya; or, The Russian
or Brave New World?” American Economic Review 109 (6): 2294–332. Nationalist Mythmaking of the Ukrainian Crisis.” Post-Soviet Affairs
Dek, Anton, Kateryna Kononova, and Tetiana Marchenko. 2019. “The Effects 32 (1): 55–74.
of Banning the Social Network VK in Ukraine.” In Rolf Fredheim et al., Lorentzen, Peter. 2014. “China’s Strategic Censorship.” American Journal of
eds., Responding to Cognitive Security Challenges. Riga: NATO Strategic Political Science 58 (2): 402–14.
Communications Centre of Excellence, 38–58. Morozov, Evgeny. 2012. The Net Delusion: The Dark Side of Internet Free-
Deutsche Welle. 2017. “Kak na Ukraine sudjat za posti v sotssetjah.” Deutsche dom. New York: Public Affairs.
Welle, June 23. https://bit.ly/34wo2v0 (accessed October 26, 2020). Nyhan, Brendan, and Jason Reifler. 2015. “Displacing Misinformation about
Dickson, Bruce J. 2016. The Dictator‘s Dilemma: The Chinese Communist Events: An Experimental Test of Causal Corrections.” Journal of Experi-
Party’s Strategy for Survival. Oxford: Oxford University Press. mental Political Science 2 (1): 81–93.
654 / Fighting Propaganda with Censorship Yevgeniy Golovchenko

Pan, Jennifer. 2017. “How Market Dynamics of Domestic and Foreign Social Stukal, Denis, Sergey Sanovich, Richard Bonneau, and Joshua A. Tucker.
Media Firms Shape Strategies of Internet Censorship.” Problems of Post- 2017. “Detecting Bots on Russian Political Twitter.” Big Data 5 (4): 310–
Communism 64 (3–4): 167–88. 24.
Peisakhin, Leonid, and Arturas Rozenas. 2018. “Electoral Effects of Biased Toal, Gerard, and John O’Loughlin. 2018. “‘Why Did MH17 Crash?’ Blame
Media: Russian Television in Ukraine.” American Journal of Political Attribution, Television News and Public Opinion in Southeastern
Science 62 (3): 535–50. Ukraine, Crimea and the De Facto States of Abkhazia, South Ossetia and
Popkova, Anna. 2019. “From Bootlegging Hollywood to Streaming Battle Transnistria.” Geopolitics 23 (4): 882–916.
Rap.” In Paolo Sigismondi, ed., World Entertainment Media: Global, Tucker, Joshua A., Andrew Guess, Pablo Barberá, Cristian Vaccari, Alexandra
Regional and Local Perspectives. New York: Routledge. Siegel, Sergey Sanovich, Denis Stukal, and Brendan Nyhan. 2018. “Social
Reisinger, Heidi, and Aleksandr Gol’c. 2014. “Hybrid War in Ukraine Russia’s Media, Political Polarization, and Political Disinformation: A Review of
Intervention and the Lessons for the NATO.” Osteuropa 64 (9–10): 119. the Scientific Literature.” Unpublished manuscript.
RFE/RL. 2017. “Poroshenko Restricts Access to Russian Websites, Social UNIAN. 2017. “‘Kievstar’ uzhe blokiruet vkontakte i odnoklassniki.” UNIAN,
Networks.” Radio Free Europe/Radio Liberty, May 16. May 17. https://bit.ly/33GPGDA (accessed October 26, 2020).
Roberts, Margaret E. 2018. Censored: Distraction and Diversion Inside Chinas Van der Vet, Freek. 2019. “Imprisoned for a ‘Like’.” In Mariëlle Wijermars
Great Firewall. Princeton, NJ: Princeton University Press. and Katja Lehtisaari, eds., Freedom of Expression in Russia’s New Media-
Robinson, Olga. 2018. “The Memes That Might Get You Jailed in Russia.” BBC sphere. Abingdon: Routledge, 209–24.
News, August 23. https://bbc.in/3boWdFp (accessed October 26, 2020). Varol, Onur, Emilio Ferrara, Clayton A. Davis, Filippo Menczer, and Ales-
Roth, Andrew. 2017. “In New Sanctions List, Ukraine Targets Russian Social- sandro Flammini. 2017. “Online Human-Bot Interactions: Detection, Esti-
Media Sites.” Washington Post, May 16. https://wapo.st/2wkQ6mZ (ac- mation, and Characterization.” In Eleventh International AAAI Conference
cessed October 26, 2020). on Web and Social Media. Palo Alto, CA: AAAI.
Sanovich, Sergey. 2017. “Computational Propaganda in Russia: The Origins Vendil Pallin, Carolina. 2017. “Internet Control through Ownership: The
of Digital Misinformation.” Working paper, New York University. Case of Russia.” Post-Soviet Affairs 33 (1): 16–33.
Shadmehr, Mehdi, and Dan Bernhardt. 2015. “State Censorship.” American VKontakte. 2017. “Kak poluchit dostup k vkontakte esli vash provayder ego
Economic Journal: Microeconomics 7 (2): 280–307. zablokiroval?” shorturl.at/gwNQZ (accessed October 26, 2020).
Shklovski, Irina, and Volker Wulf. 2018. “The Use of Private Mobile Phones at Volchek, Dmitry, and Daisy Sindelar. 2015. “One Professional Russian Troll
War: Accounts from the Donbas Conflict.” In Proceedings of the 2018 CHI Tells All.” Radio Free Europe/Radio Liberty, March 25.
Conference on Human Factors in Computing Systems. New York: Asso- Yelowitz, Aaron S. 1995. “The Medicaid Notch, Labor Supply, and Welfare
ciation for Computing Machinery, 1–13. Participation: Evidence from Eligibility Expansions.” Quarterly Journal of
Slutsky, Pavel, and Dimitrii Gavra. 2017. “The Phenomenon of Trump’s Economics 110 (4): 909–39.
Popularity in Russia: Media Analysis Perspective.” American Behavioral Zhang, Ark Fangzhou, Danielle Livneh, Ceren Budak, Lionel P. Robert Jr., and
Scientist 61 (3): 334–44. Daniel M. Romero. 2017. “Shocking the Crowd: The Effect of Censorship
Stern, Rachel E., and Jonathan Hassid. 2012. “Amplifying Silence: Uncer- Shocks on Chinese Wikipedia.” arXiv, Cornell University. https://arxiv
tainty and Control Parables in Contemporary China. Comparative Po- .org/abs/1704.00412.
litical Studies 45 (10): 1230–54. Zuckerman, Ethan. 2014. “Cute Cats to the Rescue? Participatory Media and
Stockmann, Daniela. 2013. Media Commercialization and Authoritarian Rule Political Expression.” In Danielle Allen and Jennifer Light, eds., Youth,
in China. Cambridge: Cambridge University Press. New Media and Political Participation. Cambridge, MA: MIT Press.