I introduce and study dynamic persuasion mechanisms. A principal privately observes the evolution of a stochastic process and sends messages over time to an agent. The agent takes actions in each period based on her beliefs about the state of the process and the principal wishes to influence the agent’s action. I characterize the optimal persuasion mechanism and show how to derive it in applications. I then consider the extension to multiple agents where higher-order beliefs matter.

We study torture as a mechanism for extracting information from a suspect who may or may not be informed. We show that a standard rationale for torture generates two commitment problems. First, the principal would benefit from a commitment to torture a suspect he knows to be innocent. Secondly, the principal would benefit from a commitment to limit the amount of torture faced by the guilty. We analyse a dynamic model of torture in which the credibility of these threats and promises is endogenous. We show that these commitment problems dramatically reduce the value of torture and can even render it completely ineffective. We use our model to address questions such as the effect of enhanced interrogation techniques, rights against indefinite detention, and delegation of torture to specialists.

We endogenize preferences using the “indirect evolutionary approach”. Individuals are randomly matched to play a two-person game. Individual (subjective) preferences determine their behaviour and may differ from the actual (objective) pay-offs that determine fitness. Matched individuals may observe the opponents’ preferences perfectly, not at all, or with some in-between probability. When preferences are observable, a stable outcome must be efficient. When they are not observable, a stable outcome must be a Nash equilibrium and all strict equilibria are stable. We show that, for pure-strategy outcomes, these conclusions are robust to allowing almost perfect, and almost no, observability, with the notable exception that inefficient strict equilibria may fail to be stable with any arbitrarily small degree of observability (despite being stable with no observability).

We study games of incomplete information as both the information structure and the extensive form vary. An analyst may know the payoff‐relevant data but not the players' private information, nor the extensive form that governs their play. Alternatively, a designer may be able to build a mechanism from these ingredients. We characterize all outcomes that can arise in an equilibrium of some extensive form with some information structure. We show how to specialize our main concept to capture the additional restrictions implied by extensive‐form refinements.

Many refinements of Nash equilibrium yield solution correspondences that do not have closed graph in the space of payoffs or information. This has significance for implementation theory, especially under complete information. If a planner is concerned that all equilibria of his mechanism yield a desired outcome, and entertains the possibility that players may have even the slightest uncertainty about payoffs, then the planner should insist on a solution concept with closed graph. We show that this requirement entails substantial restrictions on the set of implementable social choice rules. In particular, when preferences are strict (or more generally, hedonic), while almost any social choice function can be implemented in undominated Nash equilibrium, only monotonic social choice functions can be implemented in the closure of the undominated Nash correspondence.

We study information as an incentive device in a dynamic moral hazard framework. An agent works on a task of uncertain difficulty, modeled as the duration of required effort. The principal knows the task difficulty and provides information over time. The optimal mechanism features moving goalposts: an initial disclosure makes the agent sufficiently optimistic that the task is easy. If the task is indeed difficult, the agent is told this only after working long enough to put the difficult task within reach. The agent then completes the difficult task even though he never would have chosen to at the outset.

We introduce a class of strategies that generalizes examples constructed in two-player games under imperfect private monitoring. A sequential equilibrium is belief-free if, after every private history, each player's continuation strategy is optimal independently of his belief about his opponents' private histories. We provide a simple and sharp characterization of equilibrium payoffs using those strategies. While such strategies support a large set of payoffs, they are not rich enough to generate a folk theorem in most games besides the prisoner's dilemma, even when noise vanishes.

We study the joint design of dynamic incentives and performance feedback for an environment with a coarse (all‐or‐nothing) measure of performance, and show that hiding information from the agent can be an optimal way to motivate effort. Using a novel approach to incentive compatibility, we derive a two‐phase solution that begins with a “silent phase” where the agent is given no feedback and is asked to work non‐stop, and ends with a “full‐transparency phase” where the agent stops working as soon as a performance threshold is met. Hiding information leads to greater effort, but an ignorant agent is also more expensive to motivate. The two‐phase solution—where the agent's ignorance is fully frontloaded—stems from a “backward compounding effect” that raises the cost of hiding information as time passes.

Consider two agents who learn the value of an unknown parameter by observing a sequence of private signals. The signals are independent and identically distributed across time but not necessarily across agents. We show that when each agent's signal space is finite, the agents will commonly learn the value of the parameter, that is, that the true value of the parameter will become approximate common knowledge. The essential step in this argument is to express the expectation of one agent's signals, conditional on those of the other agent, in terms of a Markov chain. This allows us to invoke a contraction mapping principle ensuring that if one agent's signals are close to those expected under a particular value of the parameter, then that agent expects the other agent's signals to be even closer to those expected under the parameter value. In contrast, if the agents' observations come from a countably infinite signal space, then this contraction mapping property fails. We show by example that common learning can fail in this case.

We model demand for noninstrumental information, drawing on the idea that people derive entertainment utility from suspense and surprise. A period has more suspense if the variance of the next period’s beliefs is greater. A period has more surprise if the current belief is further from the last period’s belief. Under these definitions, we analyze the optimal way to reveal information over time so as to maximize expected suspense or surprise experienced by a Bayesian audience. We apply our results to the design of mystery novels, political primaries, casinos, game shows, auctions, and sports.