Bayesian Persuasion

Bayesian persuasion is a model of strategic information transmission in which a sender commits to an information structure to influence a receiver's beliefs and actions. Introduced by Kamenica and Gentzkow in 2011, it formalizes the idea that information itself can be designed, not merely transmitted. The sender chooses the signal structure that maximizes her payoff, subject to the receiver updating via Bayes' rule.

The framework applies to media bias, regulatory disclosure, and advertising — wherever a party with information advantages wants to shape decisions without controlling them directly. Bayesian persuasion is closely related to mechanism design and correlated equilibrium: all three are frameworks for shaping outcomes through information design rather than incentive design alone.

A key limitation is the commitment assumption. In many settings, senders cannot credibly commit to a signal structure. The study of strategic information transmission without commitment produces different results, revealing when persuasion is possible and when information remains trapped in silence.

Information Design and the Architecture of Belief

Bayesian persuasion is the simplest case of a broader field now called information design: the study of how structures of information — not merely the information itself — shape outcomes. The sender in Kamenica-Gentzkow does not choose what to say; she chooses what to make sayable. She designs a signal structure, a partition of the state space, a map from reality to reports. The receiver knows the map and updates accordingly. The power lies not in lying — the model forbids it — but in choosing what dimensions of reality to reveal and what to leave in shadow.

This framework illuminates domains far beyond the laboratory. In media economics, a news outlet's choice of what stories to cover and what to ignore is an exercise in Bayesian persuasion: the audience knows the outlet's editorial slant and updates beliefs about the world accordingly. In climate science communication, the IPCC's choice of confidence intervals and scenario framing shapes policy responses not by falsifying data but by structuring what decision-makers learn. In algorithmic recommendation systems, the platform's choice of what content to surface and what to suppress is a form of persuasion: the user knows the platform optimizes for engagement and rationally adjusts their beliefs about what is worth watching.

The Commitment Problem and Its Discontents

The commitment assumption — that the sender can credibly bind herself to an information structure — is the model's Achilles heel. In politics, a government that releases favorable economic data cannot commit not to have suppressed unfavorable data. In finance, a rating agency that designs a coarse rating scale cannot commit not to have calibrated the scale to maximize issuer fees. In machine learning, a platform that publishes its ranking algorithm cannot commit not to have A/B tested the algorithm against user behavior in ways that undermine the stated objective.

The study of strategic information transmission without commitment — the classic model of Crawford and Sobel — produces a very different result: information is transmitted only when the sender's and receiver's interests are sufficiently aligned, and even then it is garbled. The gap between the commitment and no-commitment models is the gap between institutional design (where law, reputation, and technology sustain commitment) and interpersonal communication (where trust must be earned in real time). Bayesian persuasion describes the world as it could be if institutions were perfect. Crawford-Sobel describes the world as it is.

Connections to Other Frameworks

Bayesian persuasion sits at the intersection of multiple traditions. It is a form of mechanism design, but one that operates on beliefs rather than incentives directly. It is a generalization of cheap talk (Crawford-Sobel), but with the added power of commitment. It is related to correlated equilibrium, in which a mediator correlates players' strategies by sending signals; Bayesian persuasion can be seen as a one-player version of this correlation problem, where the designer is the mediator and the receiver is the only player whose action matters.

In computer science, the problem of designing optimal signal structures is computationally hard: even for simple payoff structures, finding the optimal information structure is NP-hard in the number of states. This means that the elegant closed-form solutions of the Kamenica-Gentzkow model are exceptions, and real-world information design must rely on approximation, heuristics, and learning. The emerging field of algorithmic information design studies how to learn approximately optimal signal structures from data, a problem that arises naturally in online advertising, content moderation, and personalized medicine.

Bayesian persuasion is the most honest model of communication in economics — honest because it admits that information is not merely transmitted but architected. Every institution that shapes what we know — schools, media, algorithms, scientific journals — is in the business of information design. The question is not whether they persuade. The question is whether they know they are persuading, and whether we know it too. The Kamenica-Gentzkow model gives us the tools to see the architecture. Whether we have the will to look is another matter.