Rational Polarization

Abstract

Predictable polarization is everywhere: we can often predict how people’s opinions, including our own, will shift over time. Extant theories either neglect the fact that we can predict our own polarization, or explain it through irrational mechanisms. They needn’t. Empirical studies suggest that polarization is predictable when evidence is ambiguous, that is, when the rational response is not obvious. I show how Bayesians should model such ambiguity and then prove that—assuming rational updates are those which obey the value of evidence—ambiguity is necessary and sufficient for the rationality of predictable polarization. The main theoretical result is that there can be a series of such updates, each of which is individually expected to make you more accurate, but which together will predictably polarize you. Polarization results from asymmetric increases in accuracy. This mechanism is not only theoretically possible, but empirically plausible. I argue that cognitive search—searching a cognitively accessible space for a particular item—often yields asymmetrically ambiguous evidence, I present an experiment supporting its polarizing effects, and I use simulations to show how it can explain two of the core causes of polarization: confirmation bias and the group polarization effect.

Fake News!

“…In all of these cases, agents are represented as having Bayesian networks of belief (Pearl, 1986), but differing in their priors on some nodes in a way that affects how they respond to evidence directly bearing on other nodes. More striking still is recent work by Dorst (2023), who argues that higher order uncertainty regarding how to rationally respond to information can lead to polarization of a particularly strong kind, where agents can predict that they will mutually polarize. Dorst's models involve an extension of standard Bayesian representations of belief, but he offers an argument that the extension is suitable for describing rational belief revision.…”