Signals of propaganda—Detecting and estimating political influences in information spread in social networks

Social networks are a battlefield for political propaganda. Protected by the anonymity of the internet, political actors use computational propaganda to influence the masses. Their methods include the use of synchronized or individual bots, multiple accounts operated by one social media management tool, or different manipulations of search engines and social network algorithms, all aiming to promote their ideology. While computational propaganda influences modern society, it is hard to measure or detect it. Furthermore, with the recent exponential growth in large language models (L.L.M), and the growing concerns about information overload, which makes the alternative truth spheres more noisy than ever before, the complexity and magnitude of computational propaganda is also expected to increase, making their detection even harder. Propaganda in social networks is disguised as legitimate news sent from authentic users. It smartly blended real users with fake accounts. We seek here to detect efforts to manipulate the spread of information in social networks, by one of the fundamental macro-scale properties of rhetoric—repetitiveness. We use 16 data sets of a total size of 13 GB, 10 related to political topics and 6 related to non-political ones (large-scale disasters), each ranging from tens of thousands to a few million of tweets. We compare them and identify statistical and network properties that distinguish between these two types of information cascades. These features are based on both the repetition distribution of hashtags and the mentions of users, as well as the network structure. Together, they enable us to distinguish (p − value = 0.0001) between the two different classes of information cascades. In addition to constructing a bipartite graph connecting words and tweets to each cascade, we develop a quantitative measure and show how it can be used to distinguish between political and non-political discussions. Our method is indifferent to the cascade’s country of origin, language, or cultural background since it is only based on the statistical properties of repetitiveness and the word appearance in tweets bipartite network structures.