TL;DR
Misinformation isn't just about fabricated "fake news" websites; it’s about how factually true information from mainstream sources is weaponized. This paper demonstrates that articles from reliable outlets are frequently "co-shared" with fake news to support misleading narratives, effectively using the credibility of the mainstream media to grant an air of legitimacy to fringe claims.
What Problem Does This Paper Address?
Most misinformation research focuses on "source-level" analysis—identifying a "blacklist" of unreliable domains. However, these "fake news" sites actually account for a very small slice of total news consumption.
The authors argue that this binary view (Reliable vs. Unreliable) ignores a more prevalent and subtle threat: the strategic repurposing of mainstream news. A user might share a factually accurate headline from The New York Times or The Washington Post out of context to support a broader, misleading narrative (e.g., vaccine skepticism or election fraud). By focusing only on the source, researchers miss how "true" information is used to "misinform."
Methods
The researchers developed a pipeline to identify mainstream articles that act as "narrative ammunition" for misinformation circles:
- Co-sharing Analysis: They constructed a weighted graph where nodes are news URLs. An edge exists between a "fake news" URL and a "reliable" URL if they were shared by the same user. They used a graph-pruning algorithm to identify mainstream articles that are disproportionately co-shared with fake news, filtering out noise from general popularity.
- Narrative Extraction: Using an open-source NLP package called
relatio, they extracted the underlying "narrative structures" from the text. These are represented as Agent-Verb-Patient triplets (e.g.,vaccine -> cause -> shingles). - Semantic Clustering: To handle language variation (e.g., "Trump" vs. "Donald Trump"), they used K-means clustering on entity embeddings to group similar actors and objects into "low-dimensional" (broad) or "high-dimensional" (specific) narrative labels.
- Statistical Testing: They compared the frequency of these "misleading narratives" in co-shared mainstream articles versus a control group of articles from the same mainstream outlets that were not co-shared with fake news.
Datasets
- Twitter/X Panel: A massive dataset of 1.6 million U.S. users matched to their actual voter registration records (via TargetSmart), covering tweets from May 2018 to November 2021.
- NewsGuard: Used to classify domains as "fake news" (regularly publishing false content) vs. "reliable" or "trustworthy."
- Fact-Check Libraries: They aggregated ~24k unique claims labeled as "false" from sources like PolitiFact, FactCheck.org, and the Washington Post Fact Checker to build a "library" of known misleading narratives.
- Article Content: Scraped via the
Newspaper3kAPI to get full text and headlines for hundreds of thousands of URLs.
Key Insights and Ideas
- Weaponizing Credibility: Misinformation sharers prefer mainstream sources because they carry a "reputational shield." Sharing a Washington Post article is less likely to get a user flagged or socially shunned than sharing a fringe blog, even if the user is using that article to push a fringe theory.
- The "Clickbait" Trap: The study found that mainstream articles with simplified or sensationalized headlines (Type 1) are prime targets for repurposing. For example, a headline stating "Vaccinated people now make up a majority of covid deaths" is technically true (due to high vaccination rates) but is easily used to "prove" vaccines don't work.
- Temporal Repurposing: Users often dig up "old news" (Type 2) to misinform. An article from 2017 about a specific vaccine failure is reintroduced in 2021 to cast doubt on contemporary COVID-19 vaccines.
- Reach Multiplier: Co-shared mainstream articles have, on average, twice the reach (in terms of sharing users and follower counts) compared to the fake news articles they are paired with.
Key Results
- Narrative Overlap: Mainstream articles co-shared with fake news were significantly more likely to contain narratives found in misinformation content.
- Quantitative Signal: On average, 2.2% of all narratives in a co-shared mainstream article matched the "misleading narrative" library, compared to only 1.3% in the control group. While these percentages seem small, the authors note this represents a clear "signal" over the "noise" of shared language.
- Robustness: The results held true even when controlling for audience partisanship. This isn't just a "right-wing" or "left-wing" phenomenon; it’s a structural feature of how information is curated to fit worldviews.
Limitations and Future Work
- Missing User Context: The study analyzed the articles being shared but not the text of the tweets accompanying them. A user might share a mainstream article specifically to debunk it, though a manual sub-study by the authors suggested this "share-to-criticize" behavior is rare (only ~3% of cases).
- Sentence-Level Constraints: The narrative extraction tool works at the sentence level. It might miss a narrative that is built across multiple sentences or one that is negated later in the paragraph.
- Direct Quotations: While the authors did a preliminary check by removing direct quotes, they suggest future work should look deeper into whether journalists' choices of who to quote inadvertently introduces misleading narratives into mainstream reporting.
- Platform Specificity: The data is limited to Twitter/X; the dynamics might differ on more visual platforms like Instagram or closed networks like WhatsApp.
Summary updated