This paper studies how strategic information disclosure can consistently lead rational agents to abandon their initially correct model of the world in favor of a misspecified one. We study a dynamic game between a biased sender and an agent. Over an infinite horizon, the agent chooses between two ''bandit arms'' — representing alternative policies, projects, etc. — with uncertain success rates, while the sender discloses verifiable information to sway the agent towards the sender’s preferred (inferior) arm. The agent initially assumes that the sender is biased but also entertains an alternative (incorrect) model where the sender is unbiased. The agent updates their beliefs and switches models when the Bayes Factor is sufficiently high. We show how the sender can successfully mislead the agent and convince them to choose the sender-preferred arm in the long run. Moreover, we characterize when the sender can achieve this outcome with certainty.

We study learning via shared news. Each period agents receive the same quantity

and quality of first-hand information and can share it with friends. Some friends

(possibly few) share selectively, generating heterogeneous news diets across agents akin

to echo chambers. Agents are aware of selective sharing and update beliefs by Bayes’

rule. Contrary to standard learning results, we show that beliefs can diverge in this

environment leading to polarization. This requires that (i) agents hold misperceptions

(even minor) about friends’ sharing and (ii) information quality is sufficiently low.

Polarization can worsen when agents’ social connections expand. When the quantity of

first-hand information becomes large, agents can hold opposite extreme beliefs resulting

in severe polarization. We find that news aggregators can curb polarization caused by

news sharing. Our results hold without media bias or fake news, so eliminating these

is not sufficient to reduce polarization. When fake news is included, it can lead to

polarization but only through misperceived selective sharing. We apply our theory to

shed light on the evolution of public opinions about climate change in the US.

How should one incentivize creativity when acting creatively is costly? We address this question using a model of delegated bandit experimentation. A principal wants an agent to constantly switch to new arms to maximize the chance of success, while the agent faces a fixed cost of switching. We show that the principal's optimal bonus scheme is maximally uncertain---the agent receives transfers for success, but their distribution has extreme variance. Despite being stationary, this bonus scheme achieves the principal's first-best outcome. We also show that the joint surplus is strictly increasing in the agent's outside value if that value is low. To illuminate the value of opaque incentives in practice, we apply our results to the YouTube Partner Program. We argue that it uses inefficiently transparent bonuses and that better experimentation and larger profit can be achieved by using appropriate opaque bonuses.

Online voting mechanisms (e.g. polls) are a potentially powerful, cost-effective means of collecting {large} amounts of data about preferences. However, such large-scale data collection has proven to be vulnerable to sabotage (e.g. by internet trolls) if proper precautions are not taken. We consider the problem of designing a voting mechanism that is robust to derailment by external groups. We show that plurality voting and other standard mechanisms are often not robust to sabotage; in fact it is sometimes preferable to not run any poll at all. The optimal voting mechanism is found to make saboteurs indifferent between each alternative they can vote for, since this undermines their ability to adversely affect the designer's predictions of other voters' preferences.

We present a laboratory experiment designed to measure both actual and perceived dynamic inconsistency using a novel convex commitment device. Participants supply effort in the form of unpleasant tasks over time, and can commit to future effort at a cost. We find that participants demand a great deal of commitment, implying they believe they are significantly dynamically inconsistent, despite evidence of little to no dynamic inconsistency. The results suggest caution when employing commitment devices, as their usage may be unrelated to the problem they are trying to solve.