Detecting COVID-19 fake news on social media across four languages: Followers, emotions, relationships, and uncertainty

​

Fake news can kill. Many people who believed COVID-19 fake news did not get vaccinated, did not wear masks, did not social distance, got unnecessarily infected, and died (130,000 unnecessary deaths out of 217,000 [60%] US COVID-19 deaths by October 2020; Redlener et al., 2020). Furthermore, accurately detecting fake news and predicting its virality is hard. Even with training, most humans cannot identify fake news (Lutzke et al., 2019), especially as alternative media (e.g., 209 Times) can publish 99% real news mixed with 1% fake news (Shaw, 2021). Although artificial intelligence/machine learning (ML) methods can detect fake news with high accuracy (F1-ratio > 90%), they are typically atheoretical black boxes that might not apply broadly (Mohri et al., 2018). In our pilot study by contrast, we integrated situational theory of problem solving (STOPS, *Kim & Grunig, 2011) and information market theory (*Kim & Gil de Zúñiga, 2021) into a theoretical model of attributes of users (e.g., followers) and messages (personal relationship, emotion-eliciting words, vocabulary, politeness, and uncertainty markers); then we empirically showed that this model detected COVID-19 fake news with 95% accuracy (*Chiu, *Morakhovski, *Ebert et al., in press).

Also, simulation studies modeled how fake news could theoretically spread through a community (e.g., Kopp et al., 2018), but no published study has examined its actual diffusion. Without a way to effectively detect COVID-19 fake news to inform suitable interventions, many more people will die unnecessarily. 

After downloading millions of available Twitter tweets in English, Chinese, Korean, or French regarding COVID-19 from November 20, 2019 to the present, we identify tweets with links to true versus fake COVID-19 news stories (based on fact-checking sites). Then, we separately use (a) machine learning methods (support vector machine, Pisner & Schnyer, 2020; and deep neural network, Liu, 2017) and (b) statistics (statistical discourse analysis, invented by PI, *Chiu, 2008) to model fake news, before integrating them together (adding successful statistical model components into the machine learning programs). Next, we differentiate distinct online communities (via clustering algorithms, e.g., Gephi software) and model how each tweet spreads (scope and speed) both within and across online communities with another statistical method (multilevel diffusion analysis, invented by PI, Rossman, *Chiu & Mol, 2008). Overall, this project will build foundational knowledge about fake news detection that can be expanded to other languages, crises, and issues.
 

Lastly, we will improve our Social Media Analytics and Reporting Toolkit (SMART) 2.0 (Zhang, Chae, Surakitbanharn, & *Ebert, 2007) to create SMART 3.0. SMART 2.0 currently relies on human judgment of social media messages to interactively explore and analyze real-time, public Twitter and Instagram data through a user-friendly dashboard with advanced data visualizations. We will incorporate our theory, statistical models, and ML into SMART 2.0 to create SMART 3.0, a user-friendly, real time dashboard that monitors messages for likelihood of fake news, identifies current dissemination scope and speed within and across online communities, and predicts its future scope and speed. The Department of Health of Hong Kong (and of other governments) can use this dashboard (without knowing any artificial intelligence or statistics) and work with social media companies to investigate and block automated bots or groups of fake news traffickers.