Source-linked AI summary
Influence of fake news in Twitter during the 2016 US presidential election
Alexandre Bovet, Hernan A. Makse
TL;DR
The paper examines the unclear dynamics and influence of fake news on Twitter during the five months before the 2016 US presidential election. Using a large Twitter dataset, outlet classifications, network analysis, and causal modeling, it finds that fake news spreaders followed activity from Trump supporters, unlike traditional center- and left-leaning news spreaders, whose activity influenced Clinton supporters.
Problem
The study addresses limited clarity about how fake news spread on Twitter and influenced political opinion during the 2016 US presidential election.
Method
The authors analyze 171 million tweets, classify linked news outlets, characterize information-flow networks, and apply causal modeling to compare fake and traditional news diffusion.
Results
Fake and extremely biased news accounted for 10% and 15% of tweets linking to news outlets, while top fake-news spreaders followed Trump-supporter activity and traditional center- and left-leaning spreaders influenced Clinton-supporter activity.
Takeaways & Limitations
Fake news diffusion was associated with a distinct direction of influence: Trump-supporter activity influenced the dynamics of top fake-news spreaders rather than the reverse.
Takeaways & Limitations
The classification of news as fake or extremely biased is opinion-based and subject to imprecision and controversy.
Abstract
from arXiv · showhide
The dynamics and influence of fake news on Twitter during the 2016 US presidential election remains to be clarified. Here, we use a dataset of 171 million tweets in the five months preceding the election day to identify 30 million tweets, from 2.2 million users, which contain a link to news outlets. Based on a classification of news outlets curated by www.opensources.co, we find that 25% of these tweets spread either fake or extremely biased news. We characterize the networks of information flow to find the most influential spreaders of fake and traditional news and use causal modeling to uncover how fake news influenced the presidential election. We find that, while top influencers spreading traditional center and left leaning news largely influence the activity of Clinton supporters, this causality is reversed for the fake news: the activity of Trump supporters influences the dynamics of the top fake news spreaders.
1 Introduction
The 2016 US presidential election highlighted growing concern about fake news on social media and its unclear influence. This raised questions about whether online misinformation campaigns could alter public opinion and threaten election integrity.
- Fake news—fabricated or grossly distorted information—became increasingly prominent on social media during the 2016 US presidential election.The paper frames fake news as deceptive content shared through social platforms.
- The influence of misinformation in the social-media era remained unclear despite longstanding histories of misinformation and propaganda.The paper identifies massive digital misinformation as a major technological and geopolitical risk.
- Researchers questioned whether online misinformation campaigns could alter public opinion and endanger presidential-election integrity.
- The study analyzes Twitter news diffusion to investigate misinformation’s relative importance, influential spreaders, and relationship to political opinion dynamics.Its dataset covers 171 million tweets from 11 million users during the five months before election day.
2 Results
The study classifies Twitter news links, compares their diffusion networks and spreaders, and analyzes activity dynamics during the five months before the 2016 election. Fake and extremely biased news were shared by fewer but more active and densely connected users, while news dynamics separated into two correlated media clusters.
- News spreading in Twitter: 25% of tweets linked to fake or extremely biased news, although these users represented only 12% of users and posted around twice as many tweets as center- or left-news users.Fake and extremely biased news had tweet volumes comparable to center, left, and left-leaning media.
- News spreading in Twitter: Removing Breitbart, the dominant extreme-bias-right outlet with 1.8 million tweets, did not significantly change the study’s results.This robustness check addresses whether one highly active outlet drove the observed category-level patterns.
- Networks of information flow: Fake and extremely biased news networks had the highest average connectivity, about ⟨k⟩≃6.5, and were more densely connected than center and left-leaning networks.Center and left-leaning networks were larger but had more heterogeneous out-degree distributions, complicating direct diffusion comparisons.
- Top news spreaders: Top spreaders of center and left-leaning news were almost exclusively verified outlets or journalists, whereas fake-news networks included many unverified and deleted accounts.Deleted accounts were extremely active, with a median of 2,224 tweets versus 2 for users in the full dataset.
- News spreading dynamics: News activity separated into a fake, extreme-bias-right, and right cluster and a center, left-leaning, and left cluster, reflecting polarized diffusion dynamics.The separation was identified using a correlation threshold of r0 = 0.49; the authors note that structural differences may also reflect access to broadcasting technologies.
3 Discussion
Fake and extremely biased news spread through denser, more collectively driven networks than traditional news, with distinct influencer profiles and activity dynamics. The analysis suggests these categories follow different diffusion mechanisms, but its findings are limited to election-related news from popular outlets.
- 10% of news-linking tweets shared fake news and 15% shared extremely biased news; adjusting for user activity reduced their combined share to 12%.Automated accounts were similarly prevalent across categories, but bots diffusing fake news were more active.
- Fake and extremely biased news networks were denser and less heterogeneous than traditional center and left leaning news networks.Users in these networks retweeted and were retweeted by more users on average, while traditional diffusion showed more heterogeneous connectivity.
- Top traditional news spreaders were mostly verified journalists, whereas fake and extremely biased news spreaders included unverified accounts with deceptive profiles.The spreaders were identified using collective influence.
- Activity correlations formed two main media clusters, with right news outlets clustered together with fake news.The clustering was consistent with polarized communities among online news consumers.
- Fake and extremely biased news appear to diffuse through connected clusters and collective behavior rather than a small set of influencers driving cascades.Traditional center and left leaning news instead followed cascades driven by a small number of influential users in heterogeneous networks.
- The findings cannot be directly generalized to the entire Twitter population because the analysis focused on election-related news from the most popular outlets.
Methods
The study analyzes election-related Twitter activity, classifies linked news domains, identifies influential spreaders, and models causal effects in information networks. It also documents classification, domain-level, data-sharing, and methodological limitations.
- Data collection: 171 million English-language tweets mentioning Trump or Clinton were collected continuously from June 1 to November 8, 2016.
- News classification: News-linked tweets were classified by matching URL hostnames to curated fake-news, conspiracy, and bias categories, supplemented by external media-bias ratings.
- News classification: The classification was validated against Facebook-based ideological-alignment scores, yielding R2 = 0.9.
- Influence networks: Collective Influence ranks directed-network spreaders by iteratively removing nodes with the largest CIℓ,out value, using ℓ = 2.
- Influence networks: CI rankings broadly agree with high-degree and Katz rankings while also identifying locally weakly connected nodes influential at larger scales.
- Causal analysis: The study uses conditional-independence testing and causal modeling with τmax = 18 time steps, equivalent to 270 minutes, for the large multivariate dataset.
Author contributions statement
The project was conceived by H. A. M. and A. B.; A. B. performed the analysis and prepared figures, and both authors wrote the manuscript.
- H. A. M. and A. B. conceived the project, while A. B. analyzed the data and prepared figures.
- H. A. M. and A. B. jointly wrote the manuscript.
Additional information
The additional information identifies the supplementary material and the authors, and reports the authors’ competing interests.
- H. A. M. holds shares in KCore Analytics, LLC, while A. B. declares no competing interests.
- The supplementary information accompanies the paper on fake-news influence during the 2016 US presidential election.
- The authors are Alexandre Bovet and Hernán A. Makse.
Supplementary Note 1
Breitbart News was the most prominent outlet among the right-end categories by tweet volume and was closely aligned with the Trump campaign.
- 1.8 million tweets linked to Breitbart News, making it the dominant outlet among the right-end categories.
- Breitbart was closely aligned with the Trump campaign because its co-founder Steve Bannon later became the campaign’s chief executive.
Supplementary Note 2
Campaign staffers appear prominently among top news spreaders, especially on Trump’s team, but causal analysis indicates that supporter activity is driven more by top center and left-leaning spreaders than by campaign staffers.
- Campaign staff influence: Trump campaign staffers rank highly across more media categories than Clinton staffers among the top 100 news spreaders.Trump staffers appear in fake, extreme-bias right, right, right-leaning, center, and left-leaning categories; Clinton staffers appear in center, left-leaning, left, and extreme-bias left categories.
- Campaign staff influence: The Trump team played an important direct role in diffusing news on Twitter.
- Causal effects: Causal analysis finds that campaign staffers do not drive the activity of either Trump or Clinton supporters.Supporter activity is more importantly influenced by top center and left-leaning spreaders.
Supplementary Note 3
The supplementary analyses test whether outlet classification, aggregators, network structure, campaign-linked users, and automated posting alter the reported relationships among media categories and candidate supporters.
- News aggregators: Four websites are identified as at least partly aggregating news: ZeroHedge, WND, RealClearPolitics, and TruePundit.They belong to fake, extreme-bias right, right-leaning, and extreme-bias right categories, respectively.
- Classification and networks: The media categories are represented by hostname lists, tweet and user volumes, retweet-network characteristics, and correlations between category activity.Supplementary tables define the category hostnames, network statistics, and correlation profiles used in the analyses.
- Robustness analyses: Removing campaign staffers and separating Breitbart or SB+BNR into distinct categories provides supplementary causal-effect analyses.These analyses compare maximum causal effects between media spreaders and candidate-supporter activity under alternative classifications and exclusions.
- Robustness analyses: The top-100 influencer sets remain largely stable after removing news aggregators.The fake-news sets retain 96 influencers in common, as do the extreme-bias-right sets.
- Robustness analyses: The conclusions remain valid without news aggregators, with center and left-leaning influencers dominating causal effects.The supplementary analysis reports a small decrease in causal-effect intensity after removing aggregators.