[Paper Review] Prevalence of Low-Credibility Information on Twitter During the COVID-19 Outbreak

Kai‐Cheng Yang, Christopher Torres-Lugo+1 more|arXiv (Cornell University)|Apr 29, 2020

Misinformation and Its Impacts18 references100 citations

TL;DR

The paper estimates the prevalence of low-credibility information on Twitter during COVID-19, analyzes bot involvement, and identifies signs of coordinated amplification and politicization.

ABSTRACT

As the novel coronavirus spreads across the world, concerns regarding the spreading of misinformation about it are also growing. Here we estimate the prevalence of links to low-credibility information on Twitter during the outbreak, and the role of bots in spreading these links. We find that the combined volume of tweets linking to low-credibility information is comparable to the volume of New York Times articles and CDC links. Content analysis reveals a politicization of the pandemic. The majority of this content spreads via retweets. Social bots are involved in both posting and amplifying low-credibility information, although the majority of volume is generated by likely humans. Some of these accounts appear to amplify low-credibility sources in a coordinated fashion.

Motivation & Objective

• Estimate the prevalence of low-credibility information linked in tweets about COVID-19 on Twitter.
• Characterize the role of social bots in posting and amplifying low-credibility content.
• Explore signs of coordinated amplification of low-credibility information.
• Analyze topics in low-credibility content to assess politicization of the pandemic.

Proposed method

• Annotate credibility at the domain level for 570 low-credibility sources using established labeling criteria.
• Create two datasets: DS1 from a random 10% sample of tweets containing COVID-19 hashtags and links, DS2 from tweets containing low-credibility links.
• Expand shortened URLs to identify real domains and exclude Twitter and other social-media links from analysis.
• Detect bots with BotometerLite, classifying accounts as bot-like with a 0.5 threshold.
• Construct a similarity network of accounts sharing low-credibility links to identify coordinated amplification.
• Analyze article titles from DS2 to identify common low-credibility topics.

Experimental results

Research questions

• RQ1 What is the share of tweet volume that links to low-credibility sources during the COVID-19 outbreak?
• RQ2 What portion of this volume is produced or amplified by social bots?
• RQ3 Is there evidence of coordinated amplification among accounts sharing low-credibility links?
• RQ4 What topics are most prevalent in low-credibility content related to COVID-19?

Key findings

• Low-credibility links constitute 0.89% of total tweet volume in DS1, comparable to nytimes.com (0.98%) and higher than cdc.gov (<0.65%).
• About 68% of low-credibility links are shared via retweets, higher than the retweet rate for nytimes.com.
• Bot accounts contribute a higher fraction of low-credibility tweet volume than reliable sources, with bot ratios at 12.1% (low-credibility) vs 6.5% (nytimes.com) and 10.6% (overall links) at the tweet level.
• Bot participation in retweets is present, with bot-like accounts attracting more bot-like retweeters for low-credibility content than for reliable sources.
• A similarity network reveals densely connected clusters of accounts sharing many of the same low-credibility sources, suggesting coordinated amplification.
• Word cloud analysis shows politicization of the pandemic, with topics including U.S. politics, outbreak status, and economic issues; ZeroHedge is a prominent low-credibility source.