Abstract
Local news outlets play a vital role in providing trusted and relevant information to communities and addressing their specific needs and concerns. The emergence of news outlets posing as local sources and their spread on social media presents a significant challenge in the digital information landscape. This paper presents a comprehensive study investigating tweets featuring "Pink Slime" news, which is a term that has been used to refer to these news outlets due to its deceptive nature. By analyzing a large dataset of tweets, we gain valuable insights into the patterns of these tweets and the origin of these tweets including automated accounts. We show in this work that extracting syntactical features proves valuable in developing a classification approach for detecting such tweets and show that the approach achieves 92.5% accuracy. We also show that our approach achieves near-perfect detection when grouping the tweets by URL.
Introduction
In the landscape of digital misinformation, "Pink Slime" outlets represent a unique and insidious threat. Named after the low-cost beef filler, these sites—often part of large networks like Metric Media or Franklin Archer—adopt the names of small towns and cities to mimic legitimate local journalism. Because Americans generally report higher levels of trust in local news compared to national outlets, these networks serve as powerful vectors for partisan propaganda and "pay-for-play" journalism.
Despite their growing influence, there has been a lack of computational research into how this content propagates on social media. This paper addresses that gap by moving beyond static "blacklists" of domains, which are easily circumvented by new outlets, and instead focuses on identifying the unique behavioral and syntactical signatures of Pink Slime content on Twitter.
Methodology
The researchers developed a multi-step analytical pipeline to distinguish deceptive local news from legitimate reporting. The core of the approach relies on Text Segmentation, where the anatomy of a tweet is broken down by comparing it to its linked article. Tweet text is categorized into three segments:
ttl: Text copied directly from the article title.bdy: Text copied from the article body.usr: Original text added by the user.
Beyond segmentation, the study extracted 22 features across six distinct groups: structural (length, punctuation), temporal (time and day of posting), interaction (likes, retweets), and special elements (hashtags, mentions). These features were used to train a Random Forest classifier.
Data Overview
The study utilized a massive, curated dataset to ensure robust results:
- Pink Slime: 305,000 tweets containing URLs from 1,313 known Pink Slime outlets (sourced from investigations by the Tow Center and The New York Times).
- National News: 500,000 tweets from 25 outlets across the political spectrum.
- Local News: 500,000 tweets from 50 legitimate local outlets (e.g., Chicago Tribune, PennLive).
Key Insights and Analysis
The "Bot" Signature
A primary discovery was the presence of truncated sentences in Pink Slime tweets. Many of these posts copy the first sentence of an article but cut it off at an arbitrary position, even when the tweet is well within the character limit. This pattern strongly suggests automated generation by poorly configured scripts rather than human curation.
Lack of Human Engagement
Legitimate news sharing often involves personal commentary. While 32–35% of legitimate news tweets contain usr segments (original text), only 9.6% of Pink Slime tweets do. These accounts almost exclusively post "title-only" or "title + URL" content, highlighting a lack of genuine human interaction.
Temporal Surges
Pink Slime activity is not constant; it is highly cyclical and tied to political events. The researchers observed a massive surge in activity between May and September 2022, primarily driven by specific outlets targeting battleground states like Arizona, Ohio, and Illinois.
Results and Performance
The Random Forest model achieved a high level of performance, reaching 92.55% accuracy and an F1 score of 0.80 in distinguishing Pink Slime from legitimate news.
Crucially, the study found that structural and segmentation features were far more predictive than interaction metrics. This indicates that how a tweet is composed is a more reliable indicator of deception than how many people engage with it. Furthermore, the model demonstrated "zero-shot" potential, successfully identifying tweets from outlets not seen during the training phase. This suggests the model can track the emergence of new Pink Slime networks in real-time.
Finally, the researchers tested a Collective Detection strategy. By aggregating all tweets sharing a specific URL, the repetitive, mechanical nature of the sharing patterns became even more obvious. When analyzing groups of at least 100 tweets per URL, the detection accuracy reached 100%.
Conclusion and Future Work
This research provides a scalable method for detecting deceptive local news networks by focusing on the syntactical and structural patterns of their social media presence. While the current model is highly effective at identifying the posts, future work is needed to investigate the specific user accounts behind these networks to determine the balance between fully automated bots and "cyborg" (human-assisted) accounts. Additionally, applying sentiment and stance detection to the replies of these tweets could offer deeper insights into how the public perceives and interacts with this pseudo-local propaganda.
Summary updated