Source-linked AI summary
Trends in the Diffusion of Misinformation on Social Media
Hunt Allcott, Matthew Gentzkow, Chuan Yu
TL;DR
Misinformation diffusion had limited longitudinal evidence across platforms. The paper tracks interactions with 570 identified false-news sites and comparison groups from January 2015 to July 2018. Facebook interactions fell sharply after the 2016 election while Twitter interactions continued rising, though the evidence has important sampling and comparison caveats.
Problem
Evidence on whether responses to misinformation were effective and how its broader scale was evolving remained limited.
Method
The paper measures monthly Facebook engagements and Twitter shares for 570 false-news sites and comparison-site groups from January 2015 to July 2018.
Results
The Facebook-engagements-to-Twitter-shares ratio fell from roughly 40:1 to roughly 15:1 after late 2016, while comparison-site interactions remained relatively stable.
Takeaways & Limitations
The pattern is consistent with a temporary decline in the overall magnitude of misinformation and a meaningful impact from Facebook efforts after the election.
Takeaways & Limitations
The site database is likely incomplete and may overrepresent misinformation known to Facebook while excluding many small false-news sites.
Abstract
from arXiv · showhide
We measure trends in the diffusion of misinformation on Facebook and Twitter between January 2015 and July 2018. We focus on stories from 570 sites that have been identified as producers of false stories. Interactions with these sites on both Facebook and Twitter rose steadily through the end of 2016. Interactions then fell sharply on Facebook while they continued to rise on Twitter, with the ratio of Facebook engagements to Twitter shares falling by approximately 60 percent. We see no similar pattern for other news, business, or culture sites, where interactions have been relatively stable over time and have followed similar trends on the two platforms both before and after the election.
1 Introduction
The paper addresses limited evidence on how misinformation diffusion has evolved by tracking interactions with identified false-news sites across Facebook and Twitter. Interactions rose through the 2016 election, then diverged sharply across platforms, while comparison sites remained comparatively stable.
- Research scope: 570 sites identified as sources of false stories were tracked through monthly Facebook engagements and Twitter shares from January 2015 to July 2018.The study compares misinformation sites with major news, small news, and business and culture sites.
- Main trends: Interactions with false-news sites rose steadily on both platforms from early 2015 until shortly after the 2016 election.
- Main trends: More than half of Facebook interactions disappeared after the election, while Twitter interactions continued to increase.
- Main trends: The Facebook-engagements-to-Twitter-shares ratio fell from roughly 40:1 to roughly 15:1, a decline of approximately 60 percent.
- Comparison: Comparison sites remained relatively stable and followed similar Facebook and Twitter trends before and after the election.
- Remaining scale: Fake-news sites still averaged roughly 70 million Facebook engagements per month after the post-election decline.Around the election, they received almost as many Facebook engagements as the 38 major news sites in the sample.
- Caveats: The authors caution that incomplete sampling, domain-name changes, and changing demand for partisan content could partly explain the raw Facebook decline.
2 Background
The paper situates misinformation responses across platforms, legislatures, and civil society. These efforts include reducing supply, increasing friction or labeling, promoting media literacy, and supporting trust-building initiatives.
- Platform responses: Facebook and Twitter have taken steps to reduce misinformation circulation since the 2016 election.The appendix lists twelve Facebook announcements and five Twitter announcements.
- Platform responses: Platform actions broadly target misinformation supply, including blocking ads from repeat offenders and removing accounts that violate community standards.
- Government responses: Four states passed laws in 2017 encouraging media literacy and digital citizenship, while technology executives testified before congressional committees.The testimony may have raised public awareness, although there was no major national legislation.
- Civil society responses: Civil society organizations provide non-partisan media-literacy materials and fund initiatives intended to build trust between newsrooms and the public.The News Literacy Project serves teachers, while the News Integrity Initiative awarded ten grants totaling $1.8 million in 2017.
3 Data
The study constructs a database of false-news sites from five prior lists, creates comparison groups using Alexa rankings, and measures platform interactions with BuzzSumo data.
- False-news sample: Five prior research and journalistic lists were combined to identify sites producing false news stories.The lists included sources from Grinberg et al., PolitiFact, BuzzFeed, Guess et al., and FactCheck.
- Scope boundary: The database may overrepresent misinformation known to Facebook and may omit a large tail of small false-news sites.PolitiFact and FactCheck work directly with Facebook to evaluate potentially false stories flagged by users.
- False-news sample: 673 unique sites were identified, with data available for 570; robustness checks tested alternative site-selection rules.
- Comparison groups: Comparison groups were defined using Alexa category rankings and included 38 major news, 78 small news, and 54 business and culture sites.
- Interaction measures: Monthly Facebook engagements and Twitter shares for all articles were collected from BuzzSumo from January 2015 through July 2018.The measures were aggregated across sites by category and averaged by quarter.
4 Results
Fake-news interactions rose on Facebook and Twitter through the 2016 election, then diverged: Facebook fell sharply while Twitter continued rising. Other site categories showed relatively stable, similar platform trends, while fake-news interactions remained substantial and robustness checks preserved the main qualitative pattern.
- More than 50%: Facebook engagements with fake-news sites declined after the election, while Twitter shares continued to increase.Interactions had risen steadily on both platforms from early 2015 through the election.
- From around 45:1 to around 15:1: the Facebook-engagement-to-Twitter-share ratio for fake-news sites fell over two years.Ratios for major news, small news, and business and culture sites remained relatively stable.
- Roughly 70 million per month: Facebook engagements with fake-news sites still averaged this amount at the end of the sample period.Engagements fell from roughly 200 million per month at the end of 2016, but remained high.
- The main qualitative conclusions remained consistent across robustness checks using alternative site lists, activity periods, outlier exclusions, and comparison groups.The exact size and shape of the trends varied across checks.
- Figure 1 compares monthly Facebook engagements and Twitter shares across site categories, with data averaged by quarter.The figure uses BuzzSumo data and includes fake-news, major-news, small-news, and business-and-culture sites.
- Figure 2 compares the ratio of monthly Facebook engagements to Twitter shares across the site categories.The ratio is averaged by quarter using BuzzSumo data.
1 Actions Against Fake News
The paper documents platform actions aimed at reducing fake-news diffusion after the 2016 U.S. election. Facebook and Twitter actions are listed separately using announcements from their official websites.
- Appendix table 1 lists Facebook’s actions to reduce fake-news diffusion after the 2016 U.S. election.
- Appendix table 2 lists Twitter’s actions to reduce fake-news diffusion after the 2016 U.S. election.The announcements are taken from the platforms’ official websites.
2 Data
The study constructs a sample of 570 fake-news sites from five prior lists and measures their Facebook engagements and Twitter shares from January 2015 to July 2018. It compares them with major-news, small-news, and business-and-culture site categories.
- 570 sites: the final fake-news sample combines five lists after excluding 103 sites without BuzzSumo data from 673 unique sites.
- January 2015 to July 2018: the study measures total Facebook engagements and Twitter shares during this period.
- Three comparison categories: major news, small news, and business-and-culture sites were collected alongside the fake-news sample.Business-and-culture sites cover arts, business, health, recreation, and sports.
3 Robustness Checks
The robustness checks address sample-selection concerns by varying site-list inclusion, operating periods, site size, and misinformation likelihood. The downward Facebook/Twitter engagement ratio after early 2017 remains broadly consistent across these alternatives, although political demand offers another possible explanation.
- 3.1 Lists of Fake News Sites: 116 sites identified by at least two lists and 19 identified by at least three lists preserve the downward Facebook/Twitter ratio trend since early 2017.The trend is also invariant to excluding sites identified exclusively by any one of the five lists.
- 3.2 Time Coverage: The downward Facebook/Twitter ratio trend remains consistent among sites active after November 2016, active in July 2018, or active throughout August 2015–July 2018.Active operation means an Alexa global traffic rank of at least one million.
- 3.3 Number of Interactions: The downward Facebook/Twitter ratio trend survives excluding the five largest sites and appears for both the largest decile and the bottom nine deciles.This addresses concern that aggregate interactions are driven by a small number of outliers.
- 3.4 Likelihood to Publish Misinformation: Facebook engagement declines appear for black and red domains but not orange domains, with the ratio decline occurring at different times across groups.For black domains the ratio fell sharply in mid-2016 and throughout 2017; for red and orange domains, the decline mainly occurred in 2016.
- 3.4 Likelihood to Publish Misinformation: The domain-specific patterns are consistent with black and, to a lesser extent, red domains being primary targets of Facebook changes after the election.This interpretation concerns platform targeting rather than a definitive causal conclusion.
- 3.5 Sites Focusing on Political News: A cyclical decline in Facebook users’ political interest after the 2016 election could generate the falling Facebook/Twitter ratio without changes in misinformation supply or Facebook efforts.The paper tests this possibility using major political websites, whose ratio decline mainly occurred in late 2015, before the election.
Appendix Table 1: Facebook’s Actions to Fight Against Fake News
Facebook announced multiple measures against fake news, including reporting, fact-checking, ranking, contextual information, and economic-disincentive interventions. The appendix records these actions and related strategy updates through 2018.
- Appendix Table 1: Facebook’s Actions to Fight Against Fake News: Facebook announced four updates addressing hoaxes and fake news: easier reporting, disputed flags with warnings, ranking signals, and disrupted financial incentives.The measures combine user reporting, fact-checking, distribution changes, and anti-spam interventions.
- Appendix Table 1: Facebook’s Actions to Fight Against Fake News: Facebook described three broader areas of work: disrupting economic incentives, building products to curb false-news spread, and helping people make informed decisions.The table also records expansion of fact-checking, new checking techniques, repeat-offender actions, and transparency efforts.
- Appendix Table 1: Facebook’s Actions to Fight Against Fake News: Facebook tested Related Articles to give users additional information, including articles from third-party fact-checkers, when they encountered popular links.The feature was also intended to apply to potential fake-news articles.
- Appendix Table 1: Facebook’s Actions to Fight Against Fake News: Facebook later addressed cloaking, blocked ads from repeat false-news pages, replaced disputed flags with Related Articles, and prioritized friends, trustworthy publications, and local news.These updates span authenticity, monetization, context, and feed-ranking changes.
- Appendix Table 1: Facebook’s Actions to Fight Against Fake News: Facebook described a three-part strategy: remove policy-violating accounts or content, reduce false-news distribution, and provide users more context.The strategy distinguishes removal, distribution reduction, and information provision.
Appendix Table 2: Twitter’s Actions to Fight Against Fake News
Twitter’s recorded responses to misinformation emphasized content and account interventions, user reporting, bot and network research, and advertising transparency. The appendix also records the removal of fake accounts in July 2018.
- Appendix Table 2: Twitter’s Actions to Fight Against Fake News: Twitter described approaches including surfacing high-quality context, expanding resources, building tools and processes, and detecting spammy behavior at its source.These actions are presented as part of Twitter’s approach to bots and misinformation.
- Appendix Table 2: Twitter’s Actions to Fight Against Fake News: Twitter tested a feature allowing users to flag tweets containing misleading, false, or harmful information in June 2017.The feature was not officially announced.
- Appendix Table 2: Twitter’s Actions to Fight Against Fake News: Twitter shared information about malicious bots and misinformation networks potentially used in the 2016 U.S. presidential election and its efforts against them.The entry links election-related network knowledge with anti-bot and anti-misinformation work.
- Appendix Table 2: Twitter’s Actions to Fight Against Fake News: Twitter announced increased transparency for all ads on October 24, 2017, and announced removal of fake accounts on July 11, 2018.These measures address advertising transparency and account authenticity.
Appendix Table 3: 50 Largest Fake News Sites
The appendix lists the largest fake-news sites and documents the comparison groups, source classifications, activity windows, and robustness-check panels used to examine the diffusion patterns.
- Appendix Table 3: 50 Largest Fake News Sites: The table lists 50 largest fake-news sites ranked by total Facebook engagements plus Twitter shares from January 2015 to June 2018.It includes source-membership indicators and activity-period indicators for the listed sites.
- Appendix Table 3: 50 Largest Fake News Sites: The source columns identify sites from Grinberg et al., PolitiFact, BuzzFeed, Guess et al., and FactCheck, with Grinberg categories labeled black, red, and orange.A value of 1 indicates that a site appears in the corresponding source, and 0 indicates that it does not.
- Appendix Table 3: 50 Largest Fake News Sites: The comparison groups include 38 major news sites, 78 small news sites, and 54 business and culture sites selected from Alexa categories.The groups exclude government websites, databases, and sites that do not mainly produce news or other relevant content.
- Appendix Table 3: 50 Largest Fake News Sites: Robustness figures examine sites identified by at least two or three lists, exclude sites unique to each list, and vary site activity periods and size.The activity panels use Alexa global rank higher than one million to define active operation.
- Appendix Table 3: 50 Largest Fake News Sites: A separate figure examines black, red, and orange domains by monthly Facebook engagements, Twitter shares, and quarterly-average Facebook/Twitter ratios.Another figure applies the same measures to ten sites mostly focusing on political news.