Abstract
Disagreement, including on matters of fact, is a pervasive phenomenon, yet this is incompatible with existing models of social learning. I propose a new model of information processing with two key features: (i) the agent encounters false information, and (ii) the agent cannot distinguish true propositions from false ones. I study two families of axioms for update rules in this environment, finding that "willingness-to-learn" axioms are incompatible with "non-manipulability" axioms. I also provide an axiomatic characterization of update rules capturing skepticism, wishful thinking, and confirmation bias. In a simple social learning model, disagreement is not just possible, but generic. I characterize the influence of each agent on steady state beliefs and examine the impact of fact-checking.

This publication has 23 references indexed in Scilit: