Proceedings of the Annual Meeting of the Cognitive Science Society

Evidence suggests the US population has polarized over politics in recent decades, as the issue positions of America's ideological sub-groups have moved apart. I use simulations to demonstrate a plausible mechanism by which political polarization can occur among `fully' Bayesian agents. The agents learn from the testimony of competing sources using an empirically-validated Bayesian cognitive model, and do not communicate. Polarization occurs under conditions resembling real-world political debate, and the polarization produced shares several characteristics of real-world US polarization. The core of the mechanism is that agents attempt to infer and account for the bias of the information sources, but doing so creates a feedback spiral where polarized source perceptions drive polarized beliefs and vice versa, unless there is strong disambiguating evidence in one direction or another.