Our central argument is that studying the timing and centralization of message coordination helps locate the activity of a set of accounts on an empirically observable spectrum of group-based behaviors on social media. Such a spectrum ranges from “uncoordinated messages of unrelated users” to “centrally coordinated information campaign”. Astroturfing – in particular if campaigns employ unsophisticated social bots – will be placed on the latter end of the spectrum that exhibits a strong group-based coordination. One theoretically relevant question is the location of grassroots movements in this space. Grassroots movements exhibit some coordination but are less synchronized in the timing and content of their messages, because their participants respond organically to cues sent by their peers instead of centralized instructions. Therefore, they will appear in the middle of the spectrum. As a consequence, the centrally coordinated organization of astroturfing should leave different empirical traces than grassroots campaigns.
We ground our explanation of these patterns in social science theory, more specifically, the principal–agent framework as employed in political science22, where it has been used to explain the pitfalls of ground campaign organization during elections23. Applying the framework to astroturfing, we argue that the organizers of an astroturfing campaign are principals who try to pursue (political) goals by instructing and incentivizing agents to create and share messages congruent with the campaign’s goals. Because one of the purposes of astroturfing is to reach and change the behavior of as many regular users as possible, the success of the campaign is contingent on a wide reach and an organic appearance of the campaign. However, according to principal–agent theory, reaching this more complex goal is difficult because of the misalignment between the principal’s and the agents’ preferences: agents thus need to be extrinsically motivated and will try to shirk22, e.g. by creating similar or identical accounts and content instead of coming up with original contributions to the campaign. Unless the principal can establish an expensive system of close monitoring, the agents will continue to hold an information advantage over the principal: they know how much effort they exerted in creating convincing online personas (oftentimes, little), whereas the principal does not.
We thus would expect campaign accounts to post or re-post similar or identical messages within a short time window, something we call co-tweeting and co-retweeting, and coordinate with a large number of other campaign accounts. This might be similar to grassroots campaigns’ operations, but their activities do not follow centralized instructions, and are therefore more likely to post similar content in a cascading fashion over an extended time period, and with greater variation in the content as they engage in more localized coordination with a few friends, imitating and varying the content they see online. Finally, grassroots campaigns may rely heavily on Twitter’s decentralized mechanism for spreading the message, retweeting. But as astroturfing campaigns use retweeting as well10, this is unlikely to be a useful feature to distinguish the two. Finally, principal–agent theory predicts that agents will only extend their efforts when they are supervised, which might result in unique temporal activity patterns, e.g., posting only during regular office hours. We expect that this principal–agent constellation results in universal patterns that appear in astroturfing campaigns in multiple countries across the world.
Figure 1 shows a schematic representation of our research design. We use the complete data released by Twitter as part of its Information Operations Hub initiative24 up until February 2021 as “ground truth” data to show that similar coordination patterns appear in almost all cases. We also include a campaign that escaped Twitter’s attention, namely the South Korean secret service’s attempt at influencing the national elections in 2012, bringing our total population of astroturfing campaigns to 46. We then concentrate the focus of the study on the 33 campaigns that produced at least 50,000 tweets. In order to validate methods for the detection of astroturfing, we compare the behavior of astroturfing accounts to two “comparison samples” that represent groups of users that a given campaign likely tries to mimic and influence. As retreiving the tweets for such systematic samples is time and resource intensive, we select four distinct campaigns targeting audiences in six countries for an in-depth study: the Russian Internet Research Agency’s (IRA) attempt at polarizing public opinion in the U.S. and Germany, and shoring up regime support in Russia, the Chinese government’s attempt at changing the framing of the Hong Kong protest, a campaign cheerleading the government of Venezuela and the above mentioned South Korean case. Further details on case selection and sample construction can be found in the Methods section.
Defining adequate comparison groups is essential in this detection process because (decentralized) coordination can also happen as part of organic discussions among specific issue publics and grassroots movements. To demonstrate the added value of our research in that regard, we conducted a literature review of related research aiming to detect astroturfing (see Table S1 in the Supplementary Information [SI]). We excluded (1) studies that aim to detect social bots, as this research does not aim to reveal specific astroturfing campaigns but general automation patterns instead, and (2) studies that merely use the data released by Twitter for an analysis of astroturfing without aiming to predict which accounts are part of a specific campaign. The latter category includes a number of studies that examined the character and reach of individual astroturfing campaigns in various contexts such as China25, Saudia Arabia26 or the Russian influence campaign in the U.S.27,28,29,30).
Among the research listed in Table S110,31,32,33,34, the studies most closely related to ours are Vargas et al.34 and Alizadeh et al.31, both of which use some of the same data to distinguish between astroturfing accounts and regular users. In particular, Vargas and colleagues built on our earlier work10,19 with the goal of constructing a classifier to detect astroturfing campaigns more generally34. However, they chose baselines that represent neither genuine grassroots movements nor the specific issue publics targeted by the campaigns, but instead three institutionalized English-speaking elite communities (members of the U.S. Congress, the UK Parliament and academics). Not only do these communities differ from the country-specific audiences engaged in the topics targeted by an astroturfing campaign, but the U.S. and UK parliamentarians’ accounts are often run by a group of staffers who work for the politicians. These political elites’ accounts are therefore unlikely to act like an account owned by an ordinary citizen. As a second baseline, they used a snowball sample of random users without any relation to the conversations in the target countries at the time the astroturfing campaigns were active.
The paper by Alizadeh et al. focusing on English-language astroturfing campaigns uses a subset of the Twitter data to build an algorithm that can detect other astroturfing accounts in later time periods, on other platforms, or later campaigns initiated by the same actor31. In other words: their approach requires a set of astroturfing accounts already identified using alternative methods. They compare the known astroturfing accounts to a random sample of regular and politically interested users. While the latter may seem like an appropriate comparison group, the authors define politically interested users as those users who follow at least three politicians – the sample therefore may contain bots designed to boost follower counts and does not consist of accounts that engage in the debates that the astroturfing campaign tries to influence.
We construct more natural comparison samples that reflect specific contexts of each astroturfing campaign, such as country and campaign-related topics and keywords. We randomly sampled users located in the targeted country that engage in the discussions that the astroturfing campaigns try to influence to compare their activity patterns with those of astroturfing. We also take the level of activities into consideration when we design comparison samples, such that the accounts in the astroturfing campaigns and the comparison groups are comparable. Taken together, our paper presents a scalable method for detecting astroturfing campaigns and validates the findings against the very accounts the campaign tries to mimic. This universal approach does not require any training data and performs well without human inputs when detecting groups of suspicious accounts in all previously revealed instances of astroturfing on Twitter. With that, the study contributes to methodological approaches for the detection of disinformation and reveals surprising similarities in astroturfing campaigns, even across heterogeneous political and social contexts.