Social influence and network structure: How conspiracy theories spread on social media

Political conspiracy theories (CTs) pose serious risks to democratic institutions, social trust, and policymaking. While prior research has examined the psychological, cultural, or network-structural correlates of conspiracy beliefs, the social mechanisms underlying their active transmission within online networks remain insufficiently understood. This study offers empirical application of Rogers’s diffusion of innovations theory to conspiracy theory transmission, providing a dynamic account of how individuals adopt, continue, or discontinue sharing such content. Drawing on nearly 15 million tweets, retweets, replies, and quotations produced or encountered by 98 politically active Polish X (formerly Twitter) accounts over 15 months, we investigate how network exposure shape CT diffusion. Using Bayesian unordered categorical regression and network data, we examined the conditions under which these accounts acted as CT spreaders, non-spreaders, converted spreaders, and converted non-spreaders. The results indicate that the proportion of CT-spreading nodes in an account’s immediate network, rather than their absolute number, is the strongest predictor of CT transmission. Even a small fraction (1–5 %) of conspiracy theorists in one’s network significantly increases the likelihood of spreading such content. Contrary to classic diffusion theories, single-contact ties proved more influential than reinforcement from repeated-contact ties in spreading CT content. This result challenges the conventional wisdom in network diffusion theory, suggesting that even complex contagions like conspiracy sharing may spread via minimal reinforcement under certain conditions. • Fraction of CT-spreading neighbors predicts conspiracy content dissemination. • Even minimal exposure (1–5 %) greatly increases spreading likelihood. • Single-contact ties have stronger influence than repeated-contact ties in CT diffusion. • Social endorsement cues (likes) do not predict CT sharing behavior once network composition is controlled. • Uses Bayesian multinomial regression on 15 M tweets from 98 ego networks.