This study, published in early 2021, provides one of the first large-scale comparisons of how COVID-19 misinformation—the so-called "Infodemic"—propagated across Twitter and Facebook. By analyzing tens of millions of posts, the researchers reveal that the spread of low-credibility content is not merely a byproduct of automated bots, but is driven by a small, influential group of "superspreaders," many of whom hold verified status on these platforms.
TL;DR
The COVID-19 Infodemic is an "overt" phenomenon driven largely by a minority of influential, often verified, accounts rather than hidden bot networks. While misinformation trends on Twitter and Facebook are highly correlated with general pandemic interest, the specific sources and "superspreaders" vary significantly between platforms, highlighting the impact of differing user demographics and data-access policies.
What Problem Does This Paper Address?
While many studies have looked at misinformation during the pandemic, most have been siloed within a single platform. However, the modern information ecosystem is cross-platform; a narrative born on Twitter might find its greatest reach in a Facebook group.
The authors set out to bridge this gap by asking:
- How does the prevalence of low-credibility content compare between Twitter and Facebook?
- Is misinformation concentrated among a few "superspreaders"?
- Is there evidence of coordinated, inauthentic behavior (e.g., "influence operations") on both platforms?
Methods
The researchers collected data from January to October 2020 using keyword filters (coronavirus, covid, sars).
- Credibility Labeling: They employed a source-level approach, matching URLs to a list of 674 low-credibility domains (based on Media Bias/Fact Check) and 20 high-credibility benchmarks (e.g., nytimes.com, cdc.gov).
- YouTube as a Proxy: Since labeling every YouTube video is impossible, they used video unavailability (deleted or made private) as a proxy for suspicious content, validating this with manual checks on platforms like BitChute.
- Quantifying Influence: To measure how concentrated the misinformation was, they used inverse normalized entropy. A value of 1 indicates all content comes from a single source, while 0 indicates an even distribution.
- Coordination Detection: They built "co-domain networks" to identify clusters of accounts that shared suspiciously similar sets of low-credibility domains, using TF-IDF to discount the influence of very common sources.
Datasets
The study leveraged two massive, platform-specific datasets:
- Twitter: ~53 million tweets from ~12 million users, obtained via the "Decahose" (a 10% random sample of all public tweets).
- Facebook: ~37 million posts from ~140,000 public pages and groups, collected via the CrowdTangle API.
- Benchmarks: The authors curated specific lists of high-credibility sources to serve as a baseline for comparison.
Key Insights and Ideas
- The "Overt" Infodemic: Contrary to the popular image of shadowy bot armies, the primary drivers of misinformation were official handles of low-credibility sites (e.g., @gatewaypundit) and high-profile verified accounts.
- Platform Divergence: There was only a moderate correlation ($r = 0.57$) between the popularity of specific misinformation sources on Twitter versus Facebook. This suggests that "demand-side" factors—like user demographics—and "supply-side" factors—like which official handles are active on which platform—shape the misinformation landscape.
- The Verified Paradox: Verified accounts represent a tiny fraction of those sharing misinformation (~2% on Twitter, ~4.5% on Facebook), yet they are responsible for a staggering amount of the total engagement: roughly 40% of retweets on Twitter and 70% of reshares on Facebook.
- Coordination is Visible: The researchers identified clear clusters of coordinated sharing. For example, a network of Facebook pages associated with the Salem Media Group (a large U.S. radio broadcaster) was found to be a major amplifier of specific low-credibility narratives.
Key Results
- Prevalence: Twitter showed a higher ratio of low-credibility to high-credibility content (32%) compared to Facebook (21%).
- Temporal Correlation: The daily volume of misinformation on both platforms was strongly correlated ($r = 0.87$), peaking in March 2020 alongside general public attention to the pandemic.
- Engagement Ratios: The study found a massive discrepancy in how content is reshared. On Facebook, the ratio of reshares to original posts for low-credibility content was 68:1, compared to just 2.7:1 for retweets on Twitter.
Limitations and Future Work
The authors are candid about the "inconsistent data-access policies" that hamper this type of research.
- Sampling Bias: CrowdTangle only provides data for public pages and groups with large followings, missing the vast amount of misinformation potentially spreading in private Facebook profiles and smaller groups.
- Source vs. Article Level: Labeling at the domain level is a blunt instrument; a low-credibility site might occasionally publish an accurate report, and vice versa.
- Deleted Content: Because the researchers could not access posts already deleted by the platforms, their estimates of misinformation prevalence are likely "lower bounds."
Future research, they suggest, must focus on how distinct demographic groups interact with these sources and push for more transparent, privacy-sensitive data-sharing protocols between social media giants and the scientific community.
Summary updated