Political Astroturfing on Twitter: How to Coordinate a Disinformation Campaign

Abstract

Political astroturfing, a centrally coordinated disinformation campaign in which participants pretend to be ordinary citizens acting independently, has the potential to influence electoral outcomes and other forms of political behavior. Yet, it is hard to evaluate the scope and effectiveness of political astroturfing without "ground truth" information, such as the verified identity of its agents and instigators. In this paper, we study the South Korean National Information Service's (NIS) disinformation campaign during the presidential election in 2012, taking advantage of a list of campaign accounts published in court proceedings. Features that best distinguish these accounts from regular users in contemporaneously collected Twitter data are traces left by coordination among astroturfing agents, instead of the individual account characteristics typically used in related approaches such as social bot detection. We develop a methodology that exploits these distinct empirical patterns to identify additional likely astroturfing accounts and validate this detection strategy by analyzing their messages and current account status. However, an analysis relying on Twitter influence metrics shows that the known and suspect NIS accounts only had a limited impact on political social media discussions. By using the principal-agent framework to analyze one of the earliest revealed instances of political astroturfing, we improve on extant methodological approaches to detect disinformation campaigns and ground them more firmly in social science theory.

The False positive problem of automatic bot detection in social science research

“…And while this method might be more complicated and computationally expensive, it offers social scientists a way forward in identifying bots and understanding patterns of disinformation campaigns. Overall it is also important to broaden the perspective and focus more on political digital astroturfing [46] for which bots are just one potential dimension [47]. A potential solution could be to focus, as suggested in the literature, more on group behavior with unsupervised classifiers [22] in combination with a descriptive digital forensic analysis [11].…”

The False Positive Problem of Automatic Bot Detection in Social Science Research

“…Overall it is also important to broaden the perspective and focus more on political digital astroturfing (F. B. Keller, Schoch, Stier, & Yang, 2019) for which bots are just one potential dimension (Kovic, Rauchfleisch, Sele, & Caspar, 2018).…”

Behavior-Based Machine Learning Approaches to Identify State-Sponsored Trolls on Twitter

“…Furthermore, users can add URLs to their posts to direct readers to other websites that include further information that supports the contents of the tweets. Our work is based on the concept that the behavior of sponsored-trolls on social media is mainly driven by extrinsic motivations [23], this is completely different from the behavior of genuine users whose online political participation is driven by intrinsic motivations relating to self-efficacy and empowerment [24]. Although political trolls try to mimic genuine users, they exhibit suspicious patterns of behaviors because of the extrinsic motivation and because they are centrally coordinated.…”

“…The authors in [23] used principal-agent theory to explain why sponsored-trolls cannot completely hide their suspicious behaviors. This theory is usually applied in business in order to conceptualize the information asymmetry between the principal (business owner), and the agent who works in favor of the principle.…”

“…The principal can easily count the number of posts; however, assessing their qualities is difficult. Therefore, agents will create the required number of tweets without too much effort to camouflage their activities especially that usually one agent controls many accounts [23]. As a result of that, sponsoredtrolls cannot hide their behavioral pattern completely, so by selecting and extracting indicative features, we can build machine-learning classifiers that can identify these accounts based on behavior-related features.…”