Suspended accounts align with the Internet Research Agency misinformation campaign to influence the 2016 US election

TL;DR

The study analyzes a large cohort of suspended Twitter accounts to reveal alignment with the Internet Research Agency (IRA) misinformation campaign during the 2016 US election. By fusing the IRA dataset with 2016 election tweets, constructing an aggregated IRA ego network, and applying Granger causality, the authors show that suspended accounts—much more numerous than IRA accounts—had measurable influence on undecided and weak supporters, while IRA nodes themselves exerted limited direct causality. The analysis identifies two polarized communities (right and left) and a substantial suspended subset oriented toward the right, suggesting a bridging role that amplified right-leaning narratives. These findings highlight the significance of suspended accounts in shaping political discourse and call for deeper investigation into their origins, coordination, and potential collaboration with IRA-like campaigns.

Abstract

The ongoing debate surrounding the impact of the Internet Research Agency s (IRA) social media campaign during the 2016 U.S. presidential election has largely overshadowed the involvement of other actors. Our analysis brings to light a substantial group of suspended Twitter users, outnumbering the IRA user group by a factor of 60, who align with the ideologies of the IRA campaign. Our study demonstrates that this group of suspended Twitter accounts significantly influenced individuals categorized as undecided or weak supporters, potentially with the aim of swaying their opinions, as indicated by Granger causality.

Figures (5)

• Figure 1: Distribution of tweets and clients' type per account type.(a) The fraction of tweets with a URL pointing to a website belonging to one of the categories. Normalization is computed per group, meaning that, for example, the fraction of not verified tweets per category sums up to one. In each of the five groups, the order of bars is kept the same. We always display bars in the following orders: fake news, extreme bias right, right, right leaning, center, left-leaning, left, and extreme bias left. See Supplementary Table 2 for further details. (b) The fraction of tweets with a URL pointing to a website belonging to one of the categories posted through non official sources. Refer to Supplementary Table 3 for a complete view of the percentages.
• Figure 2: Share of original tweets per group. The percentage of original tweets per group and news category is presented, with normalization conducted over the total activity of each group. This means that the sum of the percentages per group represents the total fraction of original tweets. Further details can be found in Supplementary Table 6.
• Figure 3: Two-sample Kolmogorov-Smirnov tests. Results of the two sample Kolmogorov Smirnov test between groups based on their out degree. Yellow boxes indicate rejection of the null hypothesis in favor of the two-sided alternative, suggesting that the data were not drawn from the same distribution. We observe that $H_0$ is only rejected between IRA and suspended accounts, indicating a similarity in the out-degree activity of these two groups. Furthermore, verified accounts exhibit an out-degree activity that differs from that of the other groups.
• Figure 4: Account status: top 1000 active users. We show the fraction of the top 1000 active users per status (in %) in each interaction network (retweet in blue, mention in red, reply in green, and quote in gray). Possible statuses are: Verified, Suspended, Not Verified and Not Found. (a) "out" direction: users are retweeted, mentioned, quoted, or replied to by IRA users. On average, 18.4% of these users have verified accounts, 49.4% are not verified, 18.9% are suspended, and 11.4% are not found accounts (b) "in" direction: users retweet, mention, quote, or reply to IRA users. On average, 45.1% of them are not verified, 35.7% is suspended, and 17.1% is not found.
• Figure 5: Causal Networks.(a) Graph showing the maximal causal effects between the activity of the IRA nodes and the supporting classes of the presidential candidates. Arrows indicate the direction of the maximal causal effect ($\geq 0.16$) between two activity time series. The width of each arrow is proportional to the strength of the causation, and the size of each node is proportional to the auto-correlation of each time series. Dark blue and dark red highlight the contribution of strong Clinton and Trump supporters, respectively. Light blue and light red are associated with the weak Clinton and Trump supporters, gray with the undecided users, and orange with the IRA nodes. The causal relation primarily flows from strong supporters of both Trump and Clinton to weak and strong supporters of opposing political candidates. Additionally, weak supporters from both sides play a role in influencing the undecided group, with weak Trump supporters receiving support from strong Trump supporters in their efforts. Notably, IRA nodes do not play a significant role in this causal network, suggesting that they have limited influence on shaping Twitter discourse. (b) Graph showing the maximal causal effects between the activity of the suspended nodes and the supporting classes of the presidential candidates. Arrows indicate the direction of the maximal causal effect ($\geq 0.2$) between two activity time series. Strong Trump supporters have a causal effect on Suspended nodes, which, in turn, have a causal influence on both weak supporters and the undecided group. Additionally, weak supporters continue to exert a causal effect on the undecided group. Strong Trump supporters have a causal effect on strong Clinton supporters, but not vice-versa.