This dissertation studies three models of information flows in societies. In the first chapter, it considers how heterogeneity of preferences affects the accumulation of information through observational learning. In particular, it shows that arbitrarily small amount of heterogeneity can totally hinder the social learning process. The second chapter studies a situation in which an Artificial Intelligence player acts in an adversarial way against a more informed actor. It characterized the optimal learning rule for the algorithm, and quantifies the value of learning rule flexibility. The third and last chapter analyses a model of endogenous network formation, and shows that a standard rational model has very strong incentives for heterophilia, which contrasts with the echo chambers empirically observed in the literature.