Learning Through the Grapevine: The Impact of Message Mutation, Transmission Failure, and Deliberate Bias

Abstract

We analyze whether learning is simply difficult or impossible when first hand information is relayed to agents through long chains of noisy person-to-person communication. Suppose noise only takes the form of random mutations and message dependent transmission failures. Agents can still learn but at the cost of accessing many more chains of communication. They learn fully when they have more than a certain threshold number of chains and learn nothing with fewer. On a positive note, this threshold remains unchanged when agents adopt a simpler, naive learning rule. In particular, both the frequency and content of their communications are informative about the state, but agents can safely restrict attention to the more important of these dimensions. Results change dramatically when there are biased agents who relay their preferred message despite what they hear. Learning is typically impossible in the presence of even an arbitrarily small fraction of biased agents. This can explain why people become stuck at potentially different priors, despite a significant body of primary evidence pointing to one answer (e.g., disagreement on the effects of vaccination or the reality of global warming). We show that a planner can recover partial learning by limiting the number of contacts to whom agents can pass along a given message, a policy some messaging platforms are starting to use.

Keywords: Social Learning, Communication, Noise, Mutation, Bias, Fake News

JEL Classification: D83, D85, L14, O12, Z13