Comparing the Diffusion of Different Types of "fake news" and Community Responses
Overview | People | Collaborators | Sponsors | Publications | Tools
One of the challenges facing policy makers interested in determining both the impact of and appropriate interventions for the spread of "fake news" through digital social media is the variety in disinformation and misinformation sources, content, structure and targets: different actors telling different stories to different listeners who can react in different ways. The goal of this project is to determine the diffusion patterns of multiple types of disinformation within the same online conversation: who spread the stories, how fast did the stories spread, and what were the responses and how did those spread. This work hopes to inform future work aimed at developing a more detailed typology of misinformation and determining the advantages and disadvantages of different intervention strategies.
To achieve our research goal, we conducted a case study of four different types of misinformation shared in the Twitter conversation related to the release of the Black Panther Marvel superhero blockbuster, which was the most tweeted about movie ever. We collected tweet and user data (followed by/following, etc.) through Twitter's API tools using both keyword and account-based searches. The four types of misinformation we identified are as follows: Fake Attacks (claims of racially motivated violence at theaters), Satire (similar claims mocking the originals), Fake Scenes (claims the movie contained certain scenes when it did not), and Alt-Right (claims that the movie was pro-Alt-Right ideologically). Through mass media news sources and internal keyword searches we identified 304 individual tweets that fall into one of the four categories mentioned. We then compared the diffusion of these "origin" tweets by comparing when they were posted and the retweets, replies, and quotes that those tweets received over time. We also manually reviewed the replies and quotes for the most spread tweets to understand their stance (retweets are assumed to be supportive of the original tweet). We used CMU Bot-hunter, a tiered machine-learning approach, to identify automated accounts involved in the conversation. Based upon which type of disinformation/misinformation a user retweeted we separated users into different communities. We also attempted to use less imposing methods of community detection, but have experienced difficulty due to the sparsity of some of the networks involved.
One of our most interesting results to date is that it appears that parts of the Twitter community engaged in the Black Panther conversation were able to overwhelm and perhaps shutdown the original Fake Attack disinformation using a combination of Satire-type tweets and debunking retweets and replies. Satire-type stories were overwhelmingly supported and shared over the original Fake Attack stories. This is true both in terms of number of origin tweets (178 Satire vs 71 Fake Attack), number of retweets (~49,000 Satire vs ~4,600 Fake Attack) and in the types of quotes and replies (quotes and replies of Fake Attack tweets were almost entirely negative). For the most shared Fake Attack stories, quote responses that sought to debunk the original tweet were retweeted ~87,000 times, surpassing the Satire response as well. Considering the most shared Satire and Fake Attack stories, we found that Satire tweets appear to have a longer half-lives (time taken for approximately half of the total responses to be posted) than Fake Attack tweets. The half-lives of debunking responses appear to fall between those for Satire and Fake Attack tweets. We also observed that the debunking responses mostly came from either the Satire-retweeting community or from those accounts that were not found to have retweeted any of the misinformation stories.
Figure 1 # of retweets over time for the top Fake Attack tweet (green), a debunking response (red), and a Satire tweet (blue)
Figure 2 Cumulative density function for retweets of top Fake Attack, Satire, and debunking tweets