Digital Methods to Study (and Reduce) the Impact of Disinformation

Abstract

Social media have democratized communication but have led to the explosion of the so-called "fake news" phenomenon. This problem has visible implications on global security, both political (e.g.the QANON case) and health ( anti-Covid vaccination and No-Vax fake news). Models that detect the problem in real time and on large amounts of data are needed. Digital methods and text classification procedures are able to do this through predictive approaches to identify a suspect message or author. This paper aims to apply a supervised model to the study of fake news on the Twittersphere to highlight its potential and preliminary limitations. The case study is the infodemic generated on social media during the first phase of the COVID-19 emergency. The application of the supervised model involved the use of a training and testing dataset. The different preliminary steps to build the training dataset are also shown, highlighting, with a critical approach, the challenges of working with supervised algorithms. Two aspects emerge. The first is that it is important to block the sources of bad information, before the information itself. The second is that algorithms could be sources of bias. Social media companies need to be very careful about relying on automated classification.