In this thesis I explore a rational model of the emergence of mass disagreement about political issues within societies. The model supposes a recursive process whereby people attribute bias to information sources who promulgate views they disagree with, and then down-weight future information provided by those sources on the basis of their being biased. This leads people on opposing sides of political debates to attribute greater bias to their opponents than their allies, and therefore shift closer to the positions of their allies in response to subsequent debates, creating a positive feedback loop. That this process might occur is a prediction of a Bayesian Network model of how people can infer and account for source bias when applied to a simplified political information environment.

I begin by establishing the need to explain US mass belief polarisation, and suggest that the dominant existing theory – identity-motivated reasoning – has important limitations. I then explain how Bayesian agents can polarise in response to common evidence and review existing models of Bayesian polarisation. I suggest a limitation of these models is that they don’t consider source bias, and show how a model that does could lead to mass belief polarisation. In Chapter 1, I propose a Bayesian Network model of source cognition which encompasses bias, explain its behaviour, and test its descriptive validity. In Chapter 2, I test the model’s predictions about the conditions under which polarisation should occur. In Chapter 3, I analyse longitudinal data from the UK to test for a bidirectional relationship between bias perceptions and belief polarisation. I also discuss how bias perceptions might relate to affective polarisation, and test whether they do. In Chapter 4 I explore the scale of affective polarisation between factions within the same party, which provides another evaluation of the limitations of identity-motivated reasoning, and measure bias perceptions within factional contexts. In Chapter 5, I test the robustness of the claim which motivates much of this work – that belief polarisation has increased in the US, using six measures of polarisation, including a novel cluster-based method, as this is subject to some dispute. I conclude by synthesising the findings, discussing limitations, and proposing future directions.