Where Do Trolls Say They Are? Understanding the Use of Place Claims in Nation State Influence Operations

Abstract

This study employs the full public Twitter information operations archive and its nearly 90,000 accounts from at least 22 actors to examine one particular element of the strategies employed by state-backed influence actors—strategic use of place claims. Academic research exploring state-affiliated social media influence operations has, to date, focused on one or a handful of campaigns. We argue this focus has the potential to miss tactics and strategies other nations’ operations may be effectively using. Specifically, we examine account self-identified place claims to build an understanding of when and why these strategies are employed, and how these place claims relate to campaign targeting of specific populations. We examine where accounts claim to be located and how these claims relate to what we know about the targeting of each campaign. We also explore what other account behaviors are correlated with accounts’ use of specific place claims.

State-affiliated social media disinformation became an issue of international interest to academic researchers following the indictment of 13 Russian nationals for election interference in the 2016 U.S. Presidential campaign (United States of America v. Internet Research Agency LLC). These individuals were connected to the St Petersburg-based Internet Research Agency (IRA), generally believed a tool of the Russian state. In part through social media, the IRA engaged in an influence operation (IO) that spread division and discontent in the US and Western democracies more broadly (Linvill & Warren, 2020). The IRA operation included, to a large degree, the use of inauthentic social media accounts that integrated themselves into specific online communities and purported to be real people located in the US.

The use of disinformation for political purposes is not new. Octavian famously waged a successful IO campaign against Antony and Cleopatra as they vied for control of Rome in 33 B.C. (MacDonald, 2017). Social media, however, has today created a new, inexpensive, and powerful vehicle for information warfare to be waged. Social media facilitates the concealment of provenance, provides the ability to efficiently target specific populations, and speeds the rapid adaptation of tactics and messaging. Posetti & Matthews (2018) argued that social media is the greatest amplifier of disinformation since the invention of Gutenberg’s printing press. State-backed disinformation spread through social media has threatened the integrity of democratic processes, undermined trust in institutions, and fostered ideological division across the globe. The struggle against this evolving threat has been called our new “forever war” (Nunberg, 2019).

Perhaps understandably, research exploring the tactics and effects of state-affiliated social media IOs has, to date, predominately focused on campaigns originating from Russia. Researchers have examined Russian social media IOs within a single social media platform (Linvill, Boatwright, Grant, & Warren, 2019; Linvill & Warren, 2020) as well as across platforms (Dawson & Innes, 2019; Lukito, 2019). Other Russian case studies have focused on specific issues and campaigns, including the undermining of mainstream science (Strudwicke & Grant, 2020; Walter, Ophir, & Jamieson, 2020) as well as more specific issues such as the downing of MH17 over Ukraine (Vesselkov, Finley, & Vankka, 2020). Specific Russian tactics have also been analyzed, including engagement with Black Americans (Freelon et al., 2020) and the Russian IRA’s use of local news for agenda-setting purposes (Ehrett et al., 2021). Academic work examining states’ use of social media for IOs outside of the Russian context is lacking and there is inherent risk in casting Russia as a singular model of effective IO. Building too much of our understanding of state IO from the actions of a single actor will build only a limited understanding of the possible range of potentially effective tactics campaigns may engage in. While some research has examined campaigns from, for example, China (e.g., Wang, Lee, Wu, & Shen, 2020) and Iran (e.g., Al-Rawi, 2021), even most of this work has focused on how these actors have targeted the West—worse still, it often emphasizes how these other actors measure up to the “standards” of performance set by the Russians even if those standards bear no relevance to the campaign in question (Warren, Linvill, & Warren, 2023).

State-backed disinformation is not exclusively, or perhaps even primarily, a Western or Russian problem. Princeton University’s Empirical Studies in Conflict Project, which tracks ongoing state-backed IOs, has identified more than 100 influence efforts targeting 30 different countries since 2011 (Martin, Shapiro, & Ilhardt, 2020/2023, 2022). Far too little is known about most of these campaigns. Perhaps the most understudied form of social media IOs are those in which nations target their own people. Princeton’s research has identified over 30 countries that have engaged in domestic campaigns and only 10 that have targeted foreign nations. This suggests, as has been previously argued, “The global phenomenon of social media disinformation isn’t rooted in geopolitics but rather in domestic politics” (Linvill & Warren, 2021a). More research is needed to understand what tactics and techniques generalize across these campaigns, when, and why.

In this study, we focus on one small choice that must be made about each inauthentic account in each of these campaigns—where that inauthentic “troll” account publicly claims they are located. Inauthentic social media accounts can pretend to be from anywhere, with whatever level of specificity, or to make no specific place claim at all. Our goal is to better understand this one choice so we can learn something general about the tactics of state-backed IO campaigns.

We begin with a descriptive analysis characterizing the place-claiming behavior of IO accounts and exploring how these choices vary across campaigns. We structure this analysis around our first research question,

RQ1: Where (if anywhere) do social media accounts that are part of state-backed IOs purport to be located?

As we show below, there is substantial variation in place-claiming behavior, both across directing countries and across accounts operated by the same directing country. This variation suggests some patterns worth trying to explain. Specifically, we will show that many accounts make no place claim at all, but when accounts do make place claims, they seem to choose places they are targeting for influence. But that pattern leaves open the question of when and why they decide to make a claim.

In the next section, we introduce a framework that explains place-claiming behavior and derive some hypotheses from that framework. We then summarize the data we will use to evaluate this framework, and how we put those data to work. Following that, we present the results of the analyses laid out in the prior section. Finally, we discuss the implications of these results for our understanding of place-claiming, in particular, and influence operation persuasive strategies, more generally.

Framework: Place Claiming as Costly Investment

Although IO operators can choose to make any place claim whatsoever, or none at all, those choices come with costs and benefits. We can think of four potential benefits, which are logically distinct, but likely relate in practice. First, a place claim can change the set of people who are likely to follow an account, receive, and/or retweet its messages. This change may make it easier to send targeted messages or to have high confidence that the messages are reaching the intended audiences. Second, a place claim can change how the recipients of messages from the IO account interpret those messages. They may, for instance, afford messages from people that appear to be from certain places more respect, either because of a presumption of local expertise or due to some social-identity-dependent connection. A third, related, impact might be to avoid a general suspicion of accounts that hide their origin (or at least make no specific reference to it). Finally, a fourth potential benefit of a place claim is that it allows the IO actor to (falsely) represent the beliefs of people from specific places, potentially altering recipients’ perceptions about the distribution of opinions or positions of people from that place.

Most of the benefits of making a specific place claim relate to the targeting of specific audiences. Social identity theory (Israel & Tajfel, 1972) suggests that people have a tendency to perceive the world through a process of social categorization. That is, they view themselves and others primarily in terms of group membership rather than as unique individuals. The social identity model of deindividuation effects (SIDE) extends social identity theory and argues that this process is even more pronounced on social media (Reicher, Spears, & Postmes, 1995). SIDE argues that the anonymity found in computer-mediated communication shifts the relative salience of personal versus social identity to focus on the social identity. In other words, social media users are more likely to focus on what makes them part of a group rather than what makes them unique when engaging with other users.

Previous research examining state-backed social media IOs suggests that many troll accounts are deliberately engineered to resonate with the audiences being targeted (Freelon et al, 2020; Linvill & Warren, 2020; Linvill, Warren, & Moore, 2022). If the actors responsible for orchestrating such campaigns have either an intuitive or technical understanding of the processes these social identity theories help us understand, it would make sense for troll accounts to self-identify as being from the place where the audience being targeted is located. This would potentially help the troll account to appear to be a part of the same social group as the target and aid in the process of influencing the target by helping to build trust in the targeted users.

The value of each of these benefits depends on many factors but will certainly vary with the goals and overall strategic approach of the influence campaign (Linvill & Warren, 2024). Campaigns that are primarily concerned with the demotion of specific troublesome ideas, for example, may be less interested in targeted influence on specific groups. If a campaign is, for instance, simply flooding a hashtag or brigading a particular individual then place claims may have extremely limited value. On the other hand, campaigns with a goal of polarization may target their messaging to specific groups to pull them apart. Such campaigns would potentially have substantial benefits from place claims, to go along with other social-identity-based tactics. This contrast suggests a reasonable dimension for investigating the drivers of place claiming, which motivates our second research question,

RQ2: What influence goals, if any, relate to the specificity of account place claims in a state social media IO campaign?

In the context of this research question, motivated by the social-identify theory, we have one specific hypothesis we will test. Campaigns with goals that necessitate the persuasion of a target community would likely invest in individual accounts to a greater extent than campaigns that exist to simply flood hashtags or harass dissidents. So, for instance, we may be more likely to see place claims from Russian IRA accounts engaging with real Americans than we might expect from, for instance, Chinese affiliated accounts trying to disrupt the use of hashtags in conversation about the Uyghur Muslim minority (Linvill & Warren, 2021b). For these campaigns with these goals, the benefits of place names are higher, and so we expect:

H1: State social media IO campaign accounts that are part of influence efforts that involve the promotion of ideas will be more likely to make specific place claims than those in influence efforts that are primarily about the demotion of ideas.

There are also costs of making specific place claims, and they can be substantial. First, there is the additional effort required by the creators of fake accounts to identify a place and enter it into the profile information, especially if the accounts need to appear distinct. A further significant cost is that making false places claims can increase the probability of being identified as an inauthentic account. If no significant countermeasures are put in place by the IO actor, the platforms would have access to both the real and fake location of the account and could use the obvious disparity in information as an input in identifying the campaign. This would not only endanger the account itself but possibly any broader network it was a part of. Avoiding this obvious technical mismatch takes costly workarounds, like VPNs, virtual machines, or outsourcing. Relatedly, making a false place claim also raises the potential for the account operator to make some errors in their interactions that users who are familiar with the place may notice and question, negatively impacting the influence of the account.

The size of these identity-maintenance costs can also vary by campaign. Most directly, campaigns operating more accounts would have a higher cost of establishing and maintaining distinct locations and obfuscating them from detection. These costs are also likely to be higher when the false place claim is further (in some sense) from the actual location of the account operator. This contrast suggests a second useful dimension for investigating the drivers of place-claiming, which motivates our third research question:

RQ3: How, if at all, does the specificity of place claiming relate to the targeting of state-backed social-media IO campaigns?

In the context of this research question, this costs/benefits model of place claiming implies a second hypothesis about the costs of making successful specific place claims (or, relatedly, the risk of making unsuccessful ones). An IO operator who is more familiar with a place can more easily construct a false persona that can successfully pass as a resident of that place. We expect that this variation in cost would affect the attractiveness of making these claims. Therefore:

H2: The degree of specificity of state social media IO accounts’ place claims will be positively correlated with the operators’ presumed familiarity with the targeted country, its language, and its culture.

Finally, in a digital ecosystem, where bad actors simultaneously contend with not only influencing target audiences but also avoiding detection by users and platform moderation, those actors make a medley of simultaneous decisions when designing account profiles. Our final, exploratory, research question asks how these decisions move together.

RQ4: What, if any, account behaviors relate to the specificity of account place claims in a state social media IO campaign?

This last research question, RQ4, is admittedly broad. We want to explore this question to better understand what, if any, account-level features correlate with the use of place claims in an IO campaign. We can assume that actors coordinating these campaigns have made choices in how accounts are operated that they feel may further their goals. These consistencies may offer insight into the campaign as a whole. Our model of costs and benefits offers some guidance. Investing in deep personas for individual accounts is risky business. There may be contexts and times where the added effort of making a place claim simply doesn’t make sense for a given campaign. But the incentives (both costs and benefits) of those deep investments should move together. This correlation motivates our hypothesis.

H3: State social media IO campaign accounts use of specific place claims will be positively correlated with other markers of higher investment in individual accounts.

Our first two hypotheses are motivated by a change in some specific proxy for the costs or benefits of place claiming. This hypothesis, in contrast, is not motivated by such a specific change. Rather, we reason that if place-claiming is a sort of investment in persona building, any general shift in costs or benefits of such persona-building should affect all the measures of that sort of investment in the same direction, even if we cannot specify the specific changes to costs or benefits.

Data and Methods

To explore questions related to social media IO strategy, one useful resource is the Twitter IO archive. Twitter (now renamed X) maintained datasets of content from IOs they identified operating on their platform and which they attributed to the activity of specific nation states. Twitter began releasing this content in 2018 with the stated goal of enabling independent academic research and investigation. The last publicly available dataset was released in December of 2021. This repository contains more than 40 individual releases originating from 22 different nations. In the data are nearly 90,000 account records containing more than 200 million individual tweets in dozens of different languages. The data include information regarding the account profiles, the content of each message, and the engagement each message received. It also includes data addressing external links included in each message, particular hashtags employed by the campaigns, and the names of outside accounts the campaigns engaged. Important for the current study, this data also includes self-reported location data for each account as shown in the user-reported-location field.

It should be acknowledged that, for a variety of reasons, the Twitter IO archive does not offer an ideal set of data to explore this, or any, question related to state backed social media IOs. First, Twitter is only one platform of many, and what is true on Twitter may not be generalizable to other platforms. This is particularly true as Twitter use is far from uniform around the globe. Second, while they have been relatively minor, there are documented cases of Twitter making errors regarding what is included in the IO archive and which nation content has been attributed to (Calderwood, Riglin, & Vaidyanathan, 2018; Elgin, 2019). Additionally, unidentified errors likely exist. Finally, and perhaps most importantly, Twitter was never transparent about which campaigns and accounts they included in the archive or how these accounts were identified and attributed. It is possible, if not likely, that the campaigns included in the archive are not perfectly representative of the actual population of inauthentic accounts. Even if platforms suspend state-backed influence operations evenhandedly, they may not publicly attribute all state-backed campaigns evenly. Reporting suggests, for example, that both Meta and Twitter have responded to pressure from India to not attribute campaigns (Mehrotra & Menn, Nov 8, 2023; Menn & Shi, Sep 26, 2023). These limitations aside, the archive is the most robust data set currently available for academic research and it is clearly the best available dataset for analyzing our research questions.

Briefly, our approach is to rely on Twitter’s attributions concerning the actual origin of each account appearing in the publicly available IO Archive, use open-source algorithmic methods with hand-correction to infer where each account purports to be and relate the other behaviors of those accounts to their place claims to reveal what role those claims play in the strategic behavior of the inauthentic actors. For accounts that we can reasonably match in a comprehensive database of influence efforts, we supplement these data with inferences about the goals of the influence efforts to which these accounts contribute.

Original Data

Most data for this study come from the publicly available hashed version of the Twitter Information Operations archive, which includes information about more than 40 datasets of attributed platform manipulation campaigns originating from over 20 countries, spanning more than 200 million Tweets (Twitter, 2021). For each dataset, Twitter provides a short summary identifying the directing country and the main targeted countries. This archive includes nearly complete final public profile information about each account in each release, with most screennames and usernames redacted. It also contains nearly complete public tweet information, with only tweets deleted by the account before suspension was removed. The 87,437 accounts contained in these releases make up the dataset and will be the units of observation in every estimation we present. For each release, we code from the accompanying Twitter post both the originating country or countries, and the primary directing country. There are 20 unique directing countries. In order of number of sub-releases, they are: Russia (9), Iran (7), China (6), Saudi Arabia (5), Venezuela (4), Spain (2), and single releases directed by Serbia, Uganda, United Arab Emirates, Turkey, Thailand, Tanzania, Armenia, Mexico, Bangladesh, Honduras, Egypt, Ecuador, Cuba, and Indonesia.

For each account, the key field of interest is the location that the account claims, if any. This field is a free-entry text field that is fully specified by the user and is meant to be used to express where the user is located. In reality, both authentic and inauthentic accounts use this field in many different ways, including leaving it blank, making very specific place claims (“Clemson, SC, USA”), making very general place claims (“Earth” or “Somewhere”), or even filling it with nonsense terms or decorative characters.

We collect several user-defined features (Description, URL, Account Language Choice, Birth Day) and aggregate behaviors (Following Count), and build a variety of account-level metrics from the Tweet-level data for each account (Tweet Count, Retweet Count, Most Common Language, Most Common Client, Last Tweet Day), and how those tweets were engaged with (Total Retweet Count, Total Reply Count, Total Like Count). We also calculated some derivative metrics by combining some of these (Retweet Share, Maximum Life = Last Day Tweeting—Day of Birth, Tweets per Day, Total Engagement).

Secondary Data on Influence Efforts

We supplement these account-level data with secondary data on influence efforts from the Empirical Studies of Conflict Project’s “Trends in Online Influence Efforts” database (Martin, Shapiro, & Ilhardt, 2020/2023). This database is derived from secondary sources’ reporting on online influence efforts around the world. It includes extensive descriptions of the influence efforts, including whether they were active on Twitter, the originating and targeted country, the topics on which they focused, and what strategies, tactics, and goals they pursued. For each influence effort in the Martin et al database that indicates Twitter was a part of the effort, we identify the Twitter IO releases that might be related by referencing the targeted and attacking country. We then develop a set of keywords for each influence effort and search the Tweets in the indicated releases for those keywords. The list of keywords for each influence effort is included as Supplementary Appendix A.

Any account that uses the appropriate keywords is coded as contributing to that influence effort. It is possible for the same account to contribute to multiple efforts. From the database, we extract the primary goal of the influence effort, which we reduce to two categories: Promotional and demotional. Promotional goals are those that broadly include attempts to increase the level of attention received by some set of ideas and include all efforts with the primary goal coded as being Support, Spread, or Influence. Demotional goals are those that broadly include attempts to lower or limit the level of attention that some set ideas receives, includes all efforts with the original primary goal coded as Discredit or Hinder. There were also efforts with goals of Polarize, which we did not code in either group.

Place-Claim Extraction

To extract place-claim information from the user-defined text field, we employ Nominatum (Nominatum, 2022), a search API for accessing Open Street Maps (OpenStreetMaps, 2022), an open-source map of the world created by a collaboration of volunteer editors (similar to Wikipedia). We feed each user-defined-place string into the Nominatum search, and extract the country and city from each response, if there are any. For strings that are not matched to a country, we inspect each string that is used by more than one person by hand for missed matches. For strings that are matched, we inspect and correct, by hand, all matches to countries outside the top five countries for each release, as these are the matches that are most likely to be misinterpretations. This hand cleaning is important, as we made about 3000 fixes to our automated method. Most of these fixes are correcting non-claims that were misinterpreted as claims (such as “Hell” being interpreted as the small town in Michigan).

There are 24,430 accounts that we code as making country-specific place claims and, of these, we code 10,159 as making city-specific claims. There are 63,007 accounts that we code as making no country-specific claims. Of these, 61,725 simply have no text in their self-reported location field, and only 1,282 have filled that field with something that cannot be interpreted as a country claim.

For each account that makes a country-specific place claim, we further divide those claims into domestic and foreign claims. To do so, we use the Twitter release descriptions to identify the country (or countries) directing the information operation and code a place claim as domestic if the account claims to be in (one of) the country(-ies) directing the operation. We chose to code Hong Kong as “Foreign” to China, as during the majority of the time that accounts targeting it were operational, it was substantially independent.

To address RQ1, we examine these place claims and how their use varies across directing countries.

The Costs and Benefits of Place-Claiming

To evaluate whether the factors we identify in our framework relate to place-claiming, we contrast the patterns in place claims across several distinct subsets of the IO accounts.

For goals (RQ2/H1), we consider only those accounts that we could match with an influence effort in the Martin, Shapiro, & Ilhardt (2020/2023) database. Among those accounts, we simply look at differences in the share of accounts that were part of an effort with each goal class, between accounts making no place claims and those making domestic and foreign claims. For this analysis, we separate China away from the other directing countries. We do this for two reasons. First, Martin et al.’s coding is based entirely on secondary literature, reporting primarily released near the time campaigns are first identified. We argue that our understanding of these Chinese IO campaigns has evolved to be more nuanced as their tactics have been analyzed (see Linvill & Warren, 2021b). We believe that China’s influence strategy has been substantially misunderstood in early reporting as poorly executed attempts at promotion of ideas rather than what is, in fact, well executed efforts at suppressing discourse. Recent analyses of Chinese social media influence operations, including those associated with this data, suggest that campaign goals center around flooding conversations at high rates for the purpose of suppression and distraction rather than meaningful engagement and persuasion (Linvill & Warren, 2021b; Wells & Liin, 2022; Cranmer et al., 2023). This evolving understanding brings the data underlying Martin et al’s coding of China’s campaigns into question. Second, China is a massive outlier in (1) the goals (as coded), (2) place claiming, and (3) the number of accounts in the dataset. As such, it would dominate any analysis if it was included. And since we cannot do fixed-effects regressions in this setting (given the limited variance in goals within country), it is not possible to allow for China-specific effect. Instead, we present results for China, on its own, and for all other countries.

To measure the relationship between place-claims and targeting (RQ3/H2), we want to see if accounts with better understandings of the people they are targeting are more likely to make place claims. We capture these contrasts in two ways—using what Twitter says about who the campaign targets and using the most common language of the Tweets issued by the account. For the first of these categorizations, we use the disclosures accompanying each Twitter sub-release to divide the campaigns into one of two distinct sets: foreign-oriented or primarily domestic. If there was any substantial domestic component the campaign was coded as domestic. Twenty-four sub-releases, including 42,987 accounts, were coded as primarily domestic, while the remaining 23 sub-releases, including 44,450 accounts, were coded as foreign.

For the second of these categorizations, we encode whether the primary language used by the account is the most common language in the directing country. Twitter uses an algorithmic procedure to infer the language of each Tweet in the archive. For each account, primary language is coded as the most common tweet language when considering the full set of tweets from that account. An account is coded as a directing-country language match if its primary language is the most common language in the originating country. A total of 64,968 accounts are coded as directing-country language matches, while 22,469 are not.

To test the relationship between these, we run linear regressions of the form

where $Foreignid$ is a dummy variable that indicates whether account $i$ from directing country $d$ is part of a campaign that was identified by Twitter to be targeting users outside the directing country, $LanguageMatchid$ is a dummy variable indicating whether the most common language used by account $i$ from directing country $d$ is the most common language of country $d$⁠, and $Claimid$ is a dummy variable indicating whether the account makes a place claim of the indicated level of specificity (Country or City). In some specifications, we also include directing-country specific-fixed effects. Our hypotheses are that the coefficients on Foreign will be negative and those on language-match will be positive.

Correlates of Place Claims

To address our final research question (RQ4), we contrast the goals, characteristics, behavior, and engagement with accounts that make foreign and domestic place claims, both with each other and with those accounts that make no country-level place claims. We examine characteristics and behavior that we expect to relate to the costs or benefits of place claims, as outlined in H3.

For engagement, we consider a few dimensions on their own (retweets, likes, replies, follows), but also sum up all the tweet-level engagement together, transformed in logs. We do not discriminate between in-network (inauthentic) and out-of-network (authentic) engagement, so any results should be interpreted as a combination of investment in an account and the impact of that investment.

We investigate the mean (and median) differences in these characteristics, behavior, and engagement across accounts making each type of place claim, but we worry that unobservable factors about the campaigns and, especially, about the directing countries’ goals might vary in important ways that correlate with both place claims and the variables we investigate. If those factors are substantial they might obscure all-else-equal correlation amongst our variables of interest. To address this, we also estimate regression with directing-country fixed effects of the form:

where $yid$ represents the characteristic or behavior of interest for account i from directing country d, $Foreignid$ is a dummy variable representing whether the account makes a foreign place claim, $Domesticid$ for a domestic place claim, and $δd$ is a fixed-effect for each of the 20 directing countries. We interpret $βf$ as the difference between accounts making foreign place claims and those making no place claims and $βd$ as the difference between accounts making domestic place claims and those making no place claims. Of course, we are also interested in testing whether $βf=βd$⁠.

Limitations

Two elements of this study are important to keep in mind in interpreting the results.

First, they are completely dependent on the best available, but nonetheless imperfect, data on state-supported IO accounts. These data are surely a substantial subset of all state-supported IO accounts, limited to accounts identified and disclosed by Twitter during the timeframe when they were making those account-level disclosures public. To the extent that these accounts are not representative of IO accounts, more broadly, our results may not extend to that broader set.

Second, our theoretical framework of place-claiming as costly investment is post hoc, in the sense that it is informed by the initial patterns we observed in the data, rather than being fully realized and specified behind an experimental veil. A stronger test of this theory would involve collecting new data and verifying the hypotheses on those novel data. We cannot conduct that test today but look forward to doing so in future work as more data become available.

Results

RQ1 asked: Where (if anywhere) do social media accounts that are part of state IO campaigns give as their purported location? To answer this question, we used location data built from the user location field for each troll account appearing in the Twitter IO archive. Counts of the results show that, by far, the most common place trolls say they are from is nowhere: most trolls (n = 63,008, 72%) either leave the user location field blank or fill it with information that does not include a place claim that is at least as specific as a country. Table 1 shows the 10 most common cities and countries that trolls claim in their location field. Nine of the 10 most common countries appear in the Twitter IO archive as originators of campaigns (the exception being the US). Similarly, 9 of the 10 most common cities that appear in the location field of troll accounts are cities in countries that appear in the Twitter IO archive as originators of campaigns. All told 12,876 accounts (14.73%) make a domestic place claim (claim to be located in the same country as is directing the campaign) and 11,554 (13.21%) make a foreign place claim (claim to be located in a country other than that which is directing the campaign).

Figure 1 illustrates, by directing country, the percentage of accounts that make a place claim and then the percentage that make a domestic relative to foreign place claims. The figure clearly demonstrates a great deal of heterogeneity regarding place claim mix. As we see, only three countries—Armenia, Bangladesh, and Russia—have more than half of the accounts attributed to them in the IO archive that contain place claims. No country makes place claims for all of their attributed accounts. Meanwhile, China, Tanzania, and Thailand make extremely few place claims, less than 10% of attributed accounts. Finally, only one country (Armenia) exclusively makes foreign place claims.

Using Twitter attributions to discriminate between campaigns that target foreign versus domestic audiences, we see that place claiming seems to relate to targeting. For place-claiming accounts that are part of campaigns, Twitter attributes as domestically or mixed targeted, 72% of those place claims were for domestic places. For place-claiming accounts that are part of campaigns, Twitter attributes as primarily foreign targeted, only 24% of those place claims were for domestic places.

RQ2 asked: What classes of influence goals, if any, relate to account place claims in a state social media IO campaign? H1 stated that we expected accounts that had promotional goals, those related to amplification or support, would be more likely to make place claims. Table 2 shows the results of that analysis. For directing countries other than China, this hypothesis was supported by the data. For accounts with promotional goals, the share making (at least) country-level place claims was about 54%, while for those with demotional goals it was about 47% (different at p < .01). Similarly, those with promotional goals make city-level place claims about 20% of the time, compared to 18% from those with demotional goals (different at p < .01). The pattern in accounts originating from China did not fit the pattern, a result that contradicts the general hypothesis (but is consistent with our expectations, as outlined in the methods section). In the case of China, there were only 14 accounts (less than two-tenths of a percent of the China-originating accounts) coded as having demotional goals, making comparison difficult.

RQ3 asked: How, if at all, does account self-reported location relate to the targeting of state-backed social-media IO campaigns? For accounts appearing in releases that Twitter indicated were domestically or mixed targeted, 35% made a place claim. For accounts appearing in releases that Twitter indicated were foreign targeted, only 20% made a place claim. For accounts with a language match to the directing country, 27% make a place claim. For accounts which have a language mismatch to the directing country, 30% make a place claim.

Since these two factors are closely related, we evaluate H2 by considering them simultaneously. Table 3 reports the results of a multiple-regression analysis of these two factors simultaneously. In the first column, the partial correlations both move in the direction hypothesized, with accounts in foreign-targeted releases being 12 percentage point less likely to make country-level place claims (p < .01), and accounts that operate in languages that are the most common language of the directing country being seven percentage points more likely to make a place claim (p < .01). With the overall rate of place claiming being about 28%, these are substantial relationships. The second column presents the relationship between these same factors and city-specific place claims. Here, the partial correlations both also move in the direction hypothesized, with accounts in foreign-targeted releases being 3.8 percentage points less likely to make country-level place claims (p < .01), and accounts that operate in languages that are the most common language of the directing country being 6.2 percentage points more likely to make a place claim (p < .01). As the overall rate of city-place claiming is only 11.6%, these relationships are quite large.

Some of these cross-sectional relationship are robust to directing-country fixed effects and some are not. Those results are presented in the final two columns. For country-level claims, the relationships are either insignificant (foreign-targeting) or contrary to the hypothesized direction (language match, p < .01). While for city-level claiming they are smaller but still in the hypothesized direction (p < .01).

RQ4 asked: What, if any, account behaviors relate to account place claims in a state social media IO campaign? We narrowed that question a bit, through the derivation of two hypotheses, one relating to how accounts are used and one relating to the cost of building an account that could be consistent with a more specific place claim.

To evaluate H3, we present a variety of account features that relate to the level of persuasive investment, in Panel A of Tables 3 and 4. In terms of simple mean differences (Table 4) and when we adjust for directing-country fixed effects (Table 5), accounts that have made a country-specific place claim also have high levels of each of these other markers of investment. Accounts which have made a place claim tweet more and for a longer period of time and also follow more other accounts than accounts that do not make a place claim. They are more likely to include a description and a URL in their profile. Retweets also make up a relatively low share of their total output, suggesting substantial composition of original content. Perhaps in turn, accounts that have made a place claim also have greater engagement with their content (in terms of retweets, like, and replies, and the aggregate of all these and quote tweets) and more followers than accounts that have not made a place claim. Two of the markers, tweets per day and use of an unusual client, have mixed and ambiguous relationships with place claims. That fact is, perhaps, not too surprising, as in both cases higher values can represent either more or less investment. On the one hand, more tweets per day or the use of an unusual client is a marker of automation, which tends to be associated with relatively duplicative or spammy content. On the other hand, investing in automation is, itself, a marker a relatively organized campaign, which might be positively associated with quality. Across accounts from the same directing country, these markers are negatively associated with place claims, but in the cross-section they are positively associated (unusual client) or of mixed association (tweets per day).

Given the results of RQ3/H2, these the relationship could reasonably look very different between domestic and foreign place claims. In the spirit of exploration, we explore differences in behavior between accounts that make a domestic place claim relative to a foreign place claim, finding some variations worth noting. The relationship between place claims and total tweets is much stronger for domestic claims, but most other markers of investment are stronger for foreign claims, especially in the fixed-effects regressions: account age, description, and URLs, and following counts. The metrics of engagement go one way in the means, with a stronger relationship with domestic claims, but the other with directing-country fixed effects, with a stronger relationship with foreign claims.

Finally, this contrast between domestic and foreign place claims can tell us something useful about H3, that more specific place claims are more likely when I know more about a place, its language, and its culture. In Panel B of Tables 4 and 5, we present the relationship between country place claims and two other variables: whether the account operates in an language that is the most common in the directing country and whether the account makes a city-claim. First, we find that accounts making a domestic place claim are far more likely to also make that place a specific city and also more likely to operate in the language of the originating country.

Limitations

Two elements of this study are important to keep in mind in interpreting the results.

First, they are completely dependent on the best available, but nonetheless imperfect, data on state-supported IO accounts. These data are surely a substantial subset of all state-supported IO accounts, limited to accounts identified and disclosed by Twitter during the timeframe when they were making those account-level disclosures public. To the extent that these accounts are not representative of IO accounts, more broadly, our results may not extend to that broader set.

Second, our theoretical framework of place-claiming as costly investment is post hoc, in the sense that it is informed by the initial patterns we observed in the data, rather than being fully realized and specified behind an experimental veil. A stronger test of this theory would involve collecting new data and verifying the hypotheses on those novel data. We cannot conduct that test today but look forward to doing so in future work as more data becomes available.

Discussion

The results of the analysis above have substantial general implications for the strategic persuasive behavior of state-backed coordination inauthentic influence operations. We focus on a specific decision that must be made for every account in every campaign, even if the choice is just to adopt the default—namely, where the account purports to be. Broadly, we show that most accounts choose to not express a particular location claim. But when they do make such claims, they seem related to the targeting of the campaign, in the sense that domestic-targeted campaigns make mostly domestic place claims and foreign-targeted campaigns make mostly foreign claims. Motivated by this fact, we explain place-claiming behavior that considers the persuasive benefits of specific claims and the potential risks/costs. We show that elements of the campaign that make the persuasive benefits bigger (like persuasion-oriented goals) predict more and more specific place claims, while elements that make the costs of effective place claims higher (like language barriers and foreign-orientation) predict fewer and less specific place claims. We also present some suggestive evidence that place claiming is just a specific case of a broad class of investments, and that a medley of other account-specific investment behaviors move with it. As such, our cost-benefit framework might capture a more general aspect of IO campaign strategy, a general tradeoff between persuasive specificity and the risks of getting caught.

These results need to be understood in context, however. Previous research has found that most actual Twitter users fail to supply useful data in the self-identified location field. One study exploring Twitter user locations found that nearly 40% of Twitter users failed to complete this field and only 30% included meaningful information (Abbas, Bayat, & Ucan, 2018). More recent research has also found significant variation in the completion of the location field depending on the nationality of the user (Almadany, Saffer, Jameil, & Albawi, 2020). This context suggests trolls to be perhaps no less forth-coming with location information than real users, and some nation’s trolls far more so. It is possible that not making a place claim is, in fact, a choice to remain consistent with the norms of the group with which a campaign is engaging.

We have attempted to illustrate that the benefits to making a place claim are likely related to the persuasive power evoked by engaging with a persona that claims to be from a place an individual identifies with. The particular goals of a campaign clearly relate to how campaigns employ the persuasive power of place claims (RQ2). Campaigns are more likely to harness the power of place when they have promotional goals and less likely to do so with demotional goals; perhaps operating under the assumption in some cases that users are more likely to agree with and support an idea presented by a neighbor but any homeless rabble-rouser can spread division and discredit. Previous research has examined how troll accounts employ a variety of elements of identity for persuasive effect, including race (Freelon et al., 2020), ideology (Linvill & Warren, 2020), and attractiveness of profile images (Bastos, Mercea, & Goveia, 2021). Location is another, all too simple, element of the troll toolset. We also show (RQ3/H2) that place-claiming investments are risky and the risk is higher when the IO operator is targeting populations that they know less well.

Finally, this study has illustrated some compelling relationships between the use of troll account place claims and other account behaviors (RQ4). The use of place claims was correlated with other markers of investment. Attaching a place claim to an account is one of many things you can do to make an account appear more authentic and, in that way, increase the persuasive power of that account. This is consistent with one we learned from RQ2 and the relationship between place claims and campaign goals. In this same line, our analysis showed that specific city place claims are more likely when that investment is relatively easier to make (i.e., in domestic campaigns when the operator is more likely to have local knowledge to apply). Accounts are more likely to be invested in when it is seen as either necessary, easy, or both.

The heterogeneity demonstrated across campaigns and accounts analyzed in this study, even along the singular dimension of self-identified location, suggests further differences worth investigation. As previously noted, the body of literature on state-affiliated social media IOs has been dominated by empirical examinations of the Russian IRA. Expert commentary on the issue has similarly focused on Russian activity, and other actors are often evaluated simply on how well they measured up to the “Russian playbook” (Warren, Linvill, & Warren, 2023). This may convey an assumption that all IO campaigns have the same or similar goals as the Russian IRA (i.e., influence) which thereby necessitates believable personas that are capable of believable engagement with specific online communities. But what if the goal is not to influence but rather disrupt civic discourse and public dialogs?

Understanding the mind and goals of an adversary is as essential in the digital information environment as it is in any other domain. Distinct goals and contexts necessarily require distinct tactics. If this is true, then it stands to reason that there are situations and contexts where making a location claim is either not an investment worth making in a given persona or may even run counterproductive to the campaign’s goals for other extenuating circumstances. While it was the case that making a place claim often correlated with targeting, there was also a great deal of variance in whether a place claim was made at all. This study’s findings affirm the very practical reality that trolls are a tool that can be customized based on objectives and therefore should not be understood in a monolithically one-size-fits-all mindset. Future research should continue to look at state IO from a range of actors, contexts, and campaigns and build more general knowledge than has been constructed to date.

Biographical Notes

Darren L. Linvill is a Professor in the Department of Communication at Clemson University. He studies coordinated inauthentic influence operations and co-directs the Media Forensic Hub.

Jayson Warren is a PhD student in Policy Studies in the Political Science Department at Clemson University. His dissertation is about how a more holistic understanding of influence operations could inform great power competition.

Patrick L. Warren is an Associate Professor in the John E. Walker Department of Economics at Clemson University. He studies coordinated inauthentic influence operations and co-directs the Media Forensic Hub.

David L. White is a Research Professor in the Department of Parks, Recreation, and Tourism Management at Clemson University. His work includes the development of web-based metadata authoring systems, directing the deployment of hardware and software systems for real-time data collection systems, and the design and implementation of spatially enabled relational databases to support geospatial portals and web services.

Funding

This project was partially funded by support from the South Carolina Research Authority.

References