What Is a Twitter Bot?

We define Twitter bots as algorithmically controlled accounts that can automatically perform a variety of actions including posting, retweeting, liking, responding, etc. In spite of this simple definition, bots represent very diverse types of social media accounts. In some trivial cases, a bot can be as simple as a small script that redirects updates from a website to someone’s Twitter page. In more sophisticated cases, bots can fully control a Twitter account and even communicate with other social media users (Freitas et al. Reference Freitas, Benevenuto, Ghosh, Veloso, Pei, Silvestri and Tang2015), thus resembling chatbots or recomender systems one could encounter when interacting with a large company online. Here, we focus on the latter case of the more sophisticated computer programs that are capable of maintaining Twitter activity, without a continuing intervention by a human operator. Figure 1 shows an example of such a Twitter bot that adopted the old Russian male name Yermolay as its screen name.

Figure 1. Russian Twitter Bot “Yermolay”

Yermolay is a bot that is not hard to identify for a number of reasons. First, this account has posted over 12,000 tweets, but there are only six other Twitter users who follow its tweets. Second, this account does not have a user picture and employs the default silhouette image instead, which is a common sign of a bot (Nimmo Reference Nimmo2017). Additionally, the bio section for this account does not contain any meaningful information and is merely a code snippet. Finally, the tweets posted by this account resemble news headlines, another behavior typical of bots (Sanovich, Stukal, and Tucker Reference Sanovich, Stukal and Tucker2018).

Why Deploy Bots in Russia?

In recent years, fake social media accounts similar to Yermolay have gained growing popularity in Russia. Bots stand out in the repertoire of digital information manipulation tools for a number of reasons. They are hard to trace, as—unlike humans, who need to work from a physical location, which facilitates tracking their IP addresses and linking them to certain organizations like the Internet Research Agency—bots can be deployed anywhere, including cloud computing environments, digital devices, and appliances (Boulton Reference Boulton2015; Kumar and Lim Reference Kumar and Lim2020). This feature makes bots specifically suitable for covert online operations that are aimed at affecting public opinion. Bots are also automated tools that can be deployed at scale (Hegelich and Janetzko Reference Hegelich and Janetzko2016), which makes them particularly useful for specific tasks including imitating regime support by ordinary citizens or manipulating social media and search engine algorithms to promote specific content online. Last, bots are relatively cheap to employ. Unlike paid human trolls—another novel tool available to autocrats in the digital era—bots typically involve minimal human intervention and do not require salary or overtime payments. Bots are pieces of software that can run online for indefinite periods of time (until they are detected and shut down by the platforms on which they operate, or by their creators) without requiring additional resources besides the initial production costs (Agarwal Reference Agarwal2017; Pickell Reference Pickell2018).

The ease of bot deployment partially explains the utility of bots and their observed wide use in autocratic regimes. Although our knowledge of government-sponsored bot deployment initiatives is generally scarce, previous research was relatively successful in identifying the use of bots by Russian authorities at the regional level. In particular, it has been revealed that most regional governors in Russia have official Instagram accounts followed by large numbers of bots, accounting for anywhere from 13% to 63% of governors’ followers (Center for Current Politics 2019). The underlying motivation for making bots follow governors’ accounts remains unknown, but anecdotal evidence suggests two potential explanations.

First, the deployment of bots can be motivated by policy-related concerns. A recent example comes from Moscow, where the regional government employed fake social media accounts supporting a controversial housing reconstruction program on the eve of the 2018 Moscow mayoral election (Chizhova Reference Chizhova2017). Abundant evidence also suggests that pro-government bots were employed during the 2019 regional governor elections to promote positive sentiment toward incumbent candidates in a number of Russian regions (Davidov Reference Davidov2019).

There are also multiple agency-related reasons for deploying pro-government bots. On the one hand, the Kremlin has required regional governors to pay more attention to their social media activity and has advised them to create regional government departments focused on managing political communication in social media (Antonova Reference Antonova2018). On the other, public employees at these departments often lack the necessary skills for effective social media communication and rely on bots to artificially inflate relevant activity indicators. Agency-related considerations are arguably also involved in the use of bots in the interests of other levels of the Russian government. Politicians and businessmen can deploy pro-regime bots in an effort to signal loyalty or develop stronger ties to the regime, which can result in extra perks and easier access to public resources (Pertsev Reference Pertsev2019). The attractiveness of bots as a loyalty-signaling tool is due to the imbalance of public and private information in nondemocratic settings. Although private actors can use nonpublic communication channels with the government to claim responsibility for bot activity, the difficulty of tracing bots back to their masters creates relatively low risks of popular condemnation or international reputation loss for these private actors.

Our analysis in this article remains largely agnostic about specific actors coordinating or funding the bots and bot-nets we analyze, the only assumption being that those actors are interested in maximizing the intended effects of bots on other Twitter users, either for policy-related reasons or due to agency considerations. In the next section, we develop two different but related theoretical frameworks for conceptualizing the political logic behind the use of pro-government Twitter bots by autocratic regimes in their own country’s domestic politics. We focus on bots’ responses to increases in offline or online opposition activity because these periods provide bot managers with a good opportunity to signal loyalty to the regime or justify potential funding requests. Additionally, this focus allows us to link different strands of literature from the fields of comparative politics and political communication to explain a novel phenomenon in the modern politics of authoritarian regimes.

Observable Implications and Hypotheses

In this paper, we focus on four observable implications for bot behavior during offline protests and increased online opposition activity that we derive from our theoretical cost–benefit analysis framework.

One potential strategy that bots can employ is to deemphasize the protest-related agenda by distracting social media users and augmenting informational noise in their social media feeds. As a result of this strategy, if bots are activated—for whatever reason—in times of street protests or online opposition mobilization, then the volume of content posted by pro-government bots should go up during these periods. We refer to this generic observable implication as the volume implication.

Another observable implication of the same strategy is to amplify a more diverse set of news. Some of the news might be positive for the regime, others could be neutral or negative but unrelated to the cause of the protest. The goal is to distract social media users, expose them to an information environment that is rich in news of every kind, and make it harder for them to focus on the protest-related agenda (Munger et al. Reference Munger, Bonneau, Nagler and Tucker2018). We thereby expect bots to increase in the diversity of the accounts they retweet as a response to offline or online opposition mobilization, which we refer to as retweet diversity.

Another possible strategy for decreasing opposition supporters’ expected benefits is to program bots to adopt the tactics similar to what paid human trolls do in China. As King, Pan, and Roberts (Reference King, Pan and Roberts2017) show, Chinese trolls act as cheerleaders that express positive sentiment about government activities. Bots can post similar content automatically and on a large scale. More specifically, we measure this behavior with the number of tweets that pro-government bots post about Vladimir Putin on a given day. We refer to this response as cheerleading.

Finally, the automated trolling and harassment of opposition leaders constitutes an alternative strategy aimed at increasing the expected costs of supporting opposition. We measure the use of this strategy, referred to as negative campaigning, with the number of pro-government bot-produced tweets that mention Alexey Navalny, a prominent Russian opposition leader, who is also known for his charisma and ability to bring large numbers of protesters to the streets (Nechepurenko Reference Nechepurenko2018).

These four observable implications provide us with a set of hypotheses that can be tested empirically to both shed light on the logic behind the use of bots in domestic politics in modern nondemocratic regimes and contrast the use of bots during mass street protests and their deployment as a response to opposition mobilization online. Are bots primarily used as yet another tool for demobilizing citizens in times when opposition is trying to bring people onto the streets, or are they mainly employed as an online agenda control, or gatekeeping, mechanism (McCombs and Shaw Reference McCombs and Shaw1972) tailored to regulating information flows on social media?

In Table 1, we concisely summarize our hypotheses that are drawn from the two different—but not mutually exclusive—theoretical frameworks for explaining the use of pro-regime political bots in comparative authoritarian regimes. The first is that bots are used in the interest of offline demobilization—that is, to reduce participation in offline protests. The second is that bots are used to control the online agenda, which we refer to as the online agenda control framework, and therefore will be mobilized in response to opposition online activity.Footnote ¹ These two theoretical perspectives do not differ in the hypothesized tactics that will be used by pro-government bots but only in the events (offline protest vs. online activity) that will trigger the use of these bots. Thus, in Table 1, the four observable implications that start with H1 (H1a, H1b, H1c, and H1d) are derived from the offline demobilization theoretical framework, whereas the four implications that start with H2 (H2a, H2b, H2c, and H2d) are the predicted empirical observations from the online agenda control theoretical framework.

Table 1. Summary of Hypotheses

Note: The observable implications that start with H1 (H1a, H1b, H1c, and H1d) together make up the tests of the offline demobilization theoretical framework, whereas the implications that start with H2 (H2a, H2b, H2c, and H2d) make up the tests of the online agenda control theoretical framework.

Detecting Bots

As social media bots are gaining growing attention in the social sciences, tools for detecting them are continually developing.Footnote ² Most existing research relies on publicly available general-purpose software like Botometer (Heredia, Prusa, and Khoshgoftaar Reference Heredia, Prusa and Khoshgoftaar2018; Stieglitz et al. Reference Stieglitz, Brachten, Ross and Jung2017; Vosoughi, Roy, and Aral Reference Vosoughi, Roy and Aral2018) or DeBot (Bello, Heckel, and Minku Reference Bello, Heckel and Minku2018; Schuchard et al. Reference Schuchard, Crooks, Stefanidis, Croitoru, Aiello, Cherifi, Cherifi, Lambiotte, Lió and Rocha2019) that are not tailored to the specific contexts under study here. This is particularly problematic, as bots can exhibit dramatic differences in patterns across national borders, periods, and campaigns (Uyheng and Carley Reference Uyheng, Carley, Thomson, Bisgin, Dancy, Hyder and Hussain2019). Additionally, the use of tools such as Botometer or DeBot does not allow researchers to retrospectively analyze historical data—both approaches rely on a real-time call to the Twitter API, which attempts to assess the bot versus human status of the account at the time of the study. For many studies we need to classify bots at the time of the account activity in question—which is of course crucial for many research tasks in the social sciences.Footnote ³

Another approach involves researchers using leaked lists of bots, as has been previously applied when studying paid human trolls (Sobolev Reference Sobolev2018). However, this technique requires both the existence of such a leak and confidence in its veracity and representativeness.

To avoid the concerns with the aforementioned approaches and to enable better quantification of error, we make use of a tool that is designed specifically for detecting bots on the Russian political Twitter and shows high out-of-sample performance (Stukal et al. Reference Stukal, Sanovich, Bonneau and Tucker2017). This approach to bot detection relies on a machine learning algorithm that is trained on human-labeled data of Russian Twitter accounts and enables researchers to scale up human coding of Russian bots through training an ensemble of supervised classifiers. This method can also be trained and implemented for distinct periods as needed.

More specifically, this bot detection algorithm involves four main steps. First, it takes 10 random samples of the accounts that were annotated by a group of trained human coders who labeled the accounts as bots or humans. Then, for each of the sampled subsets of the data, it uses a variety of features measured at the account level to predict the assigned labels via training four different supervised classifiers (a ridge-penalized logistic regression, a support vector machine, an extreme gradient boosted tree, and an adaptive boosting binary classifier). Third, it uses a unanimous voting rule to aggregate the results of the four classifiers for each subset of the data. And fourth, it aggregates the results across all the subsets by applying a majority voting rule.

As discussed in much more detail in Stukal et al. (Reference Stukal, Sanovich, Bonneau and Tucker2017), the architecture of this bot detection tool is specifically designed to maximize precision over recall—that is, to minimize the probability that a human account would be predicted to be a bot (for more details, please refer to Online Appendix D). From a substantive perspective, this implies that we might miss some of the most sophisticated bots in this study, especially the ones that could deceive human coders. On the other hand, Stukal et al. (Reference Stukal, Sanovich, Bonneau and Tucker2017) have shown that the relatively unsophisticated bots detected by this method make up around half of all Twitter accounts that regularly tweeted about Russian politics in Russian during 2014–2016. Therefore, the types of bots we are able to study here comprise not only a sizeable portion of all bots but also a decent share of all Russian-language Twitter accounts—bots and humans—that discuss Russian politics. An additional reason for us for being conservative with bot detection is to ensure that the activity we analyze in this article is in fact automated and our findings properly characterize the usage of a large set of bots in Russia, thereby shedding new light on the information manipulation strategies available to modern nondemocratic regimes. Nevertheless, the limitations of the bot-detection algorithm we employ in this article also highlight the importance of continued research in this area, as the type of bots we identify and track during this period may or may not be the same types of bots that are prevalent in the 2020s.

Measuring the Political Orientation of Bots

Identifying Russian Twitter bots is not sufficient for studying the strategies of their deployment in the interests of the government in that previous research has shown that pro-government bots are not the only type of Twitter bots that are active in Russia. In fact, according to Stukal et al. (Reference Stukal, Sanovich, Bonneau and Tucker2019b), only 27% of bots operating in Russian political Twitter during this period were pro-government; the remaining 73% were either neutral (49%) or appeared to be friendly to regime opponents (24%).

For this reason, we moved beyond mere bot detection and employed another tool specifically designed to measure Russian bots’ political orientation via a multilayer feed-forward neural network (for details on this classifier, please refer to Online Appendix E). This orientation classifier demonstrated high performance in separating pro-government and antigovernment bots (Stukal et al. Reference Stukal, Sanovich, Bonneau and Tucker2019a).Footnote ⁴ As our primary interest lies in understanding how Russian pro-government actors use bots during periods of increased opposition activity (either offline or online), we focus on 1,516 pro-government Twitter bots (with over one million tweets in our collection in total) detected with high confidence in Stukal et al. (Reference Stukal, Sanovich, Bonneau and Tucker2019a). Although this is only a subset of all the detected pro-government bots in Russia, we focus on these accounts, as we can be confident that they are bots tweeting pro-Kremlin messages and that this set is large enough to inform us about patterns of bot activity.

Identifying Offline Protests

Since 2011, when the largest political protests in Russia under Putin started, there have been a number of important and prolific academic efforts to track and describe Russian protests (Gabowitsch Reference Gabowitsch2016; Greene Reference Greene2014; Lankina and Tertytchnaya Reference Lankina and Tertytchnaya2019; Lankina and Voznaya Reference Lankina and Voznaya2015; Robertson Reference Robertson2011; Volkov Reference Volkov and Ross2015). However, studying protest in almost real time is still a challenging research task that requires real-time data analytics and automated event detection technologies that are still open areas of research in computer science (Hasan, Orgun, and Schwitter Reference Hasan, Orgun and Schwitter2018). We make use of the database of the Integrated Crisis Early Warning System (ICEWS) project that automatically extracts information about events from news articles based on the pattern “who did what to whom, when, and where” (Boschee et al. Reference Boschee, Lautenschlager, O’Brien, Shellman, Starz and Ward2015). Different types of events have different codes in the database, including a variety of codes for subtypes of protests, which allows researchers to extract potential protest-related data for different countries without substantial time gaps. We extract information on all protests in Russia during 2015–2018 from the ICEWS database and manually validate the results by searching for any mentions of protest events on a given date in Russian and English-language mass media. We also augment the ICEWS database with additional protest events that it clearly missed. Finally, we restrict our attention to protests with at least 1,000 participants, as we believe these events are large enough that the Russian government would be motivated to take online action against them. The full list of protest rallies in Russia we analyze is reported in the Online Appendix I.

A potential limitation of the ICEWS data is its reliance on media reports. This might be problematic for studying protest in government-controlled media systems like Russia, as media might be required to underreport political protests. To address this concern, we conduct a set of robustness checks using an extended version of the protest data set that includes the augmented ICEWS data, data from Lankina and Tertytchnaya (Reference Lankina and Tertytchnaya2019), and the Mass Mobilization in Autocracies Database (Weidmann and Rød Reference Weidmann and Rød2019). We find substantively similar results (see Online Appendix J).

Identifying Spikes in Online Opposition Activity

To identify periods of increased online opposition activity, we count the total number of tweets that prominent opposition leaders or related organizations posted on their Twitter accounts. Specifically, we collected data on 15 Twitter accounts that belong to activists, independent journalists, or mass media with extensive and benevolent coverage of the Russian opposition (Table 2).

Table 2. Opposition and Independent Journalists

To identify spikes, we use a robust measure that compares daily numbers of tweets from these 15 accounts with the respective median over a two-month period. More specifically, we define a spike as a day with at least five times as many tweets from opposition accounts as they posted on a median day a month before and after that day. This measurement instrument identifies 24 days with spikes in the opposition tweeting activity. As shown in the Online Appendix G, only two of these spikes coincide with days of offline protests.

Measuring Effects

To measure the effect of protest rallies and spikes in opposition activity on bot strategies, we use maximum likelihood to estimate a hierarchical varying-intercept varying-slope Poisson regression model with the following parameterization of the conditional mean function:

$$ {\displaystyle \begin{array}{l}\mathcal{E}\left({Y}_{it}|{D}_t,{\gamma}_{m\left[t\right]}\right)\\ {}\hskip1em ={\displaystyle \begin{array}{l}\exp \Big({\beta}_{0i}+{\beta}_{1i}\times \mathrm{Protes}{\mathrm{t}}_t+{\beta}_{2i}\times \mathrm{Onlin}{\mathrm{e}}_t+{\beta}_{3i}\\ {}\times \mathrm{Placeb}{\mathrm{o}}_t+{\gamma}_{m\left[t\right]}\Big),\end{array}}\end{array}} $$

where $ {Y}_{it} $ is a count variable for the number of tweets pro-government bot i posted on day t; Protest _t , Online _t , and Placebo _t are binary independent variables that equal 1 if there is a street protest, spike in the online opposition activity on day t, or day t is a randomly selected placebo day, respectively; and $ {\gamma}_{m\left[t\right]} $ are month-year random effects.

This flexible parameterization (for alternatives, please refer to the Online Appendix N) allows us to overcome the limitation of constant-slope models that can fail to capture data heterogeneity by constraining regression coefficients to be the same across different groups of observations (Gelman and Hill Reference Gelman and Hill2007).

For the ease of interpretation, we focus on SAPD, simulated average predictive differences (Gelman and Hill Reference Gelman and Hill2007, Chapter 5.7), defined as follows:

(1) $$ {\displaystyle \begin{array}{rcl} SAPD& =& \frac{1}{S}\sum \limits_{s=1}^S\frac{1}{M\times I}\sum \limits_{m=1}^M\sum \limits_{i=1}^I\Big(\unicode{x1D53C}\left({Y}_{it}|{D}_t=1,{\gamma}_{m\left[t\right]}{\hat{\beta}}^{(s)}\right)\\ {}& & -\unicode{x1D53C}\left({Y}_{it}|{D}_t=0,{\gamma}_{m\left[t\right]},{\hat{\beta}}^{(s)}\right)\Big),\qquad \end{array}} $$

where I is the total number of pro-government bots, M is the total number of year-months, s is the index for a simulated coefficient vector, and S is the total number of simulations.

SAPD is a discrete analogue of the simulated average marginal effects that shows the average change in the expected value of the dependent variable as the value of a binary independent variable shifts from zero to one. We make 1,000 draws from the sampling distribution of coefficient estimators (S = 1,000). Then, for each simulated coefficient vector, each pro-government bot, and each year-month, we compute the difference in the expected values of the dependent variable if there is a protest or online spike on a given day or if there is none and we average these differences across bots and month-years. Finally, we average the computed average differences over all simulated coefficients.

Finally, to accommodate the fact that we perform a large number of statistical tests in this study and control the probability of Type I errors, we employ the Bonferroni correction for 191 tests. In other words, we use the nominal confidence level of $ 1-\frac{\alpha }{191} $ , where $ \alpha \hskip1.5pt =\hskip1.5pt 0.05 $ .

Causality

The causal attribution of the observed SAPD to protest rallies or spikes in online opposition activity requires making certain assumptions. First, it requires that bots could change their behavior as a result of an intervention (e.g., via an intervention of a bot manager). We call this the manipulability assumption. This assumption would be violated if bots were preprogrammed to exhibit no temporal volatility. In that case, their tweeting activity would be constant over time. Our tweet volume dimension provides a simple test for this assumption, as it shows whether bots change the intensity of their tweeting during specific events.

Alternatively, this assumption would be violated if bots were preprogrammed to react to activists’ tweets.Footnote ⁵ In this case, bots and activists would show similar—if not identical—patterns of activity. However, as we show in Figure 2 (Section 1), a fivefold increase in the number of activists’ tweets per day (the online spikes dummy goes from 0 to 1) is on average associated with only a one-tweet increase in the activity of bots. Thus, there is little evidence of bots being preprogrammed to react to activists.

Figure 2. Volume of Tweets: Effect Size

Another assumption needed to causally interpret SAPD requires that the effect of protest rallies or spikes in online activity be separable from the effects of other events on the ground. We refer to this as the identifiability assumption. Although it is challenging to fully verify this assumption, we address the issue in the Online Appendix G by applying a structural topic model to the texts of the tweets and identifying topics that stand out as prevalent on days of offline protests and show that all these topics are related to protests or protest causes. This finding supports the validity of the identifiability assumption.

In addition, as we show in Online Appendix G, the effects of street protests are separable from the effects of spikes in the online activity of the Russian opposition, as only two of these spikes overlap with street protests.