The COVID-19 Infodemic: Twitter versus Facebook

Contributions

1. First large-scale comparative study of COVID-19 misinformation prevalence across Twitter and Facebook, analyzing over 53M tweets and 37M posts
2. Measured concentration of low-credibility content sharing among root accounts and identified "superspreaders" using inverse normalized entropy
3. Documented role of verified accounts in amplifying low-credibility information (40% of retweets on Twitter, 70% of reshares on Facebook)
4. Identified clusters of coordinated accounts sharing suspiciously similar sets of low-credibility domains, with evidence of coordinated amplification
5. Provided evidence for bot-assisted amplification of misinformation on Twitter using BotometerLite scoring

Method

The authors collected COVID-19-related content from both platforms using keywords ("coronavirus," "COVID," "sars") from January 1 to October 31, 2020. Low-credibility domains were identified by matching URLs against the Media Bias/Fact Check (MBFC) website, which catalogs 674 low-credibility domains by reporting accuracy level. High-credibility sources were drawn from a curated list of 20 major news outlets (e.g., BBC, New York Times, Reuters).

Twitter data: Decahose (10% random sample) from the Decahose API, capturing over 53 million tweets posted by about 12 million users. Data collection provided access to original tweets, retweets, accounts, and user networks.

Facebook data: Posts/search API endpoint from CrowdTangle, yielding over 37 million posts from public pages and groups. For each post, aggregate statistics (reshares, comments, reactions) were recorded but individual user-level retweet information was not available.

YouTube data: Queried the YouTube API to identify videos linked on both platforms; focus was on suspicious videos identified through list of deleted or private videos.

Measurement of concentration: Inverse normalized entropy formula to quantify whether low-credibility sharing is distributed across many users or concentrated among a few accounts. Defined "superspreaders" as root accounts disproportionately responsible for distributing low-credibility content.

Network analysis: Constructed co-domain networks weighted by term frequency-inverse document frequency (TF-IDF) similarity to identify clusters of accounts sharing suspiciously similar low-credibility domain combinations. Applied thresholds for minimum link counts and similarity scores to filter for suspicious patterns.

Bot detection: Used BotometerLite (rapidscore.com/OSoMe/botometer-pro) to score Twitter accounts for bot-like behavior (0–1 scale).

Results

Prevalence of low-credibility content:

• Twitter: 245,620 original tweets + 653,415 retweets linking to low-credibility domains (32% of original tweets overall)
• Facebook: 303,119 original posts + 20.46M reshares linking to low-credibility domains (21% of original posts overall)
• Low-credibility content peaked during March 2020 when COVID-19 cases surged globally, then declined to a stable baseline by summer

Superspreaders:

• Concentration of influence is highly skewed on both platforms: small minority of accounts responsible for majority of low-credibility sharing
• Verified accounts receive disproportionately more retweets (40% on Twitter, 70% on Facebook) despite comprising only 1.9% of root accounts on Twitter and 4.5% on Facebook
• Of the 23 top low-credibility sources, 19% are verified on Twitter and 21% are verified on Facebook — lower than general verified-account prevalence
• Verified accounts tend to share links to low-credibility content that receive significantly higher engagement

Platform differences:

• Sources rank differently in popularity across platforms (Spearman r = 0.57, p < 0.01)
• Right-wing and state-affiliated media sources dominate low-credibility sharing on both platforms
• YouTube video prevalence differs: links to YouTube appear with higher frequency on both platforms but suspicious videos show different rank distributions

Coordinated behavior:

• Identified clusters of accounts sharing suspiciously similar combinations of low-credibility domains
• On Twitter, identified high-confidence coordinated clusters; Facebook clusters less clear due to data limitations
• Suggests coordinated amplification by both authentic organizations and potentially inauthentic accounts

Bot amplification:

• Sources with more bot activity on Twitter are equally shared on both platforms (weak bot amplification after controlling for platform differences)

Notes

Strong methodological contribution through careful platform comparison while acknowledging data-collection biases (Twitter biased toward active users, Facebook toward high-popularity content). The finding that verified accounts play an outsized role despite being underrepresented among low-credibility sources is particularly important for platform governance — verification status appears to amplify reach regardless of content quality. The coordinated-behavior findings are suggestive but constrained by Facebook data limitations. The paper acknowledges that deleted tweets/posts and platform deletions of misinformation may underestimate true infodemic prevalence.