Exposure to the Russian Internet Research Agency foreign influence campaign on Twitter in the 2016 US election and its relationship to attitudes and voting behavior

The study addresses whether exposure to Russia’s Internet Research Agency (IRA) foreign influence campaign on Twitter during the 2016 US presidential election was associated with changes in Americans’ political attitudes, polarization, and voting behavior. The context is a shift from initial optimism about social media’s democratic potential to concerns that states use these platforms for domestic control and foreign election interference. While evidence documents the IRA’s activities and suggests goals of aiding Donald Trump and exacerbating polarization, data linking individual-level exposure to behavioral outcomes have been scarce. Prior work offers inconsistent measures of exposure and lacks pre-election, longitudinal linkage to users’ feeds. The authors leverage survey-linked Twitter data to quantify exposure’s scale and distribution, identify who was exposed, and estimate whether exposure correlated with within-person changes in salient policy attitudes, perceived candidate polarization, and vote choice. The importance lies in assessing the real-world impact and limits of foreign influence efforts on social media during a high-salience election.

The paper situates itself within research on foreign influence campaigns and social media, noting that prior studies examined interactions with IRA accounts and broader campaign structures/content but lacked adequate temporal coverage, often used data collected after platforms removed accounts, and provided only short windows of observation post-election. Political science literature generally finds minimal persuasive effects from traditional campaign communications, even under favorable conditions. Research on misinformation/fake news shows exposure is concentrated among small groups and strong partisans, raising the possibility that foreign influence exposure may likewise be concentrated among those least susceptible to persuasion. Platform-released aggregate figures (e.g., Facebook’s and Twitter’s non-comparable metrics) further complicate understanding of reach. This study addresses these gaps by linking pre- and during-campaign survey measures to contemporaneous Twitter timelines.

The authors conducted a three-wave longitudinal survey of 1,496 US Twitter users (fielded by YouGov) during the 2016 presidential campaign (Wave 1 in April 2016; final pre-election wave late October 2016; post-election recontact for vote). Respondents consented to provide their Twitter account information. Researchers collected the list of accounts each respondent followed and retrieved all tweets from those followed accounts during the eight months before Election Day, yielding approximately 1.2 billion posts in aggregate. Using Twitter’s Election Integrity data releases, the authors identified tweets originating from state-backed foreign influence accounts (primarily the Russian IRA; also China, Iran, Venezuela). Exposure was operationalized as potential exposure: tweets and retweets from foreign influence accounts that appeared in a respondent’s timeline via followed accounts (direct or, predominantly, via retweets/quote tweets). They quantified exposure volume and distribution over time and compared it to exposure from US national news media and US politicians during the campaign’s final month. For inferential analyses, they exploited the panel structure to estimate within-respondent changes. Outcomes included: (1) changes in positions on eight salient policy issues and overall ideology (scales recoded so higher values align with Trump positions), (2) changes in perceived candidate polarization (distance between respondents’ placements of Trump and Clinton on the same issue and ideological scales), and (3) changes in vote preferences and vote choice. Vote outcomes compared first-wave candidate preference rankings to final pre-election rankings and to validated/post-election reported vote, and a broader measure capturing shifts that could benefit either candidate via third-party voting or abstention. Primary models were OLS regressions predicting within-respondent changes from exposure measured as log(exposures + 1) and as a binary indicator (any exposure). Controls included socio-demographics, social media use, and partisanship where appropriate; alternative specifications included Poisson models and models predicting final outcomes conditional on baseline values. They conducted equivalence testing using Two One-Sided Tests (TOST) to bound effect sizes, and performed back-of-the-envelope national exposure estimates by combining sample-based exposure rates with Pew estimates of Twitter penetration (21% of US adults in 2016) and US Census adult population figures.

- Scale and concentration: 70% (n=1,042) of respondents were exposed to at least one post from a foreign influence campaign between April–November 2016, with 786,634 such posts identified across timelines. Daily total exposures (aggregated across respondents) ranged from roughly 2,000 early in the campaign to ~10,000 at its height, peaking at ~24,000 on Election Day. Russian accounts comprised 86% of all foreign-campaign exposures. Exposure occurred mainly via retweets, rising to 75–80% by Election Day. - Exposure concentration among users: 1% of respondents accounted for 70% of exposures to Russian foreign influence posts; 10% accounted for 98%. By contrast, for domestic content, 1% of respondents accounted for 24% of exposures to news media and 37% to politicians, indicating foreign influence exposure was particularly concentrated. On the supply side, 1% of Russian accounts generated 89% of the content appearing in timelines. - Relative scale versus domestic content: In the final month, respondents were exposed on average to ~4 Russian foreign influence posts per day, versus ~106 posts from national news media and ~35 posts from US politicians. Median weekly exposure to Russian accounts was zero in each week of the final month. - Partisanship and exposure: Self-identified Strong Republicans saw roughly nine times as many Russian foreign influence posts as Democrats or Independents. Regression models confirmed a monotonic increase in exposure with Republican identification; no evidence of a U-shaped pattern (i.e., equally high exposure among Strong Democrats). - Aggregate US exposure estimate: Combining sample-based exposure rates (63% of US Twitter users exposed to at least one Russian post) with Pew (21% US adult Twitter penetration) and Census data (244.8M US adults), the authors estimate ~51M US Twitter users in 2016 and approximately 32 million potentially exposed to IRA content over the eight months pre-election. - Attitudes and polarization: Across nine ideological/issue outcomes and perceived polarization measures, associations with exposure were near zero and statistically insignificant; effect signs were not in the pro-Trump direction. Equivalence tests indicate standardized effects around 0.05 SD (issues) and 0.06 SD (polarization), and for all but one of 18 estimates, effects greater than 0.2 SD can be rejected. - Voting behavior: Linear probability models show near-zero, statistically insignificant relationships between exposure and (a) switching from initial preference to vote for Trump, (b) switching in pre-election preferences between Trump and Clinton, or (c) broader shifts benefiting one candidate via third-party voting or abstention. Coefficient signs were negative, inconsistent with pro-Trump effects.

The findings indicate that while foreign influence content circulated widely in absolute terms, exposure was highly concentrated among a small subset of users—especially strong Republicans already predisposed toward Trump—and was overshadowed by domestic political content from news media and politicians. Within-person analyses show no meaningful relationship between exposure and changes in salient policy attitudes, perceived candidate polarization, or voting behavior. This pattern addresses the core research question: despite substantial public concern and the campaign’s documented activities, the observed exposure on Twitter during 2016 does not translate into detectable shifts in attitudes or vote among ordinary users. The results align with broader evidence of minimal persuasive effects of political messaging and suggest that the IRA’s social media efforts likely had limited capacity to change preferences or behavior, given the audience composition and the larger information environment. The authors note potential second-order effects—such as heightened public perceptions of interference and resulting mistrust—that are not captured by individual-level attitude and voting measures.

The paper contributes by combining longitudinal survey data with contemporaneous Twitter timelines to directly quantify exposure to a major foreign influence campaign and to estimate its association with changes in attitudes and behavior. It documents substantial absolute exposure but extreme concentration among strong Republicans and a relative scale dwarfed by domestic political content. Across multiple outcomes, evidence indicates no meaningful relationship between exposure and shifts in issue positions, perceived polarization, or vote choice. These results temper claims that such campaigns can easily manipulate ordinary social media users. Future research should examine other platforms and media formats (images, videos), potential cross-platform dynamics, and second-order effects (e.g., trust in elections). Improved access to platform data, refined measures of actual attention rather than potential exposure, and analyses in different political contexts are important to assess whether effects could emerge under different conditions or evolving tactics.

- Exposure measurement is potential exposure (tweets appearing in timelines via followed accounts), not verified viewing; actual attention cannot be observed. - Reliance on Twitter’s identification of foreign influence accounts; the classification process is not publicly verifiable by the authors. - Observational design limits causal inference; exposure is not randomly assigned, and selection into following retweeters or content affinities may confound relationships. - The study focuses on Twitter and text-based posts; it does not assess other platforms (e.g., Facebook, YouTube), other content types (images, videos), or non-social media interference channels (e.g., hacking/phishing). - Exposure is heavily concentrated among strong Republicans; results may reflect this audience composition and may not generalize to different distributions of exposure or future contexts.