📈 economics

Revealed Bayesian Persuasion

This paper provides necessary and sufficient conditions for an analyst to empirically test whether a decision maker's observed state-dependent choices are consistent with being influenced by an unobserved sender optimizing their expected payoff through Bayesian persuasion.

Original authors: Jeffrey Mensch

Published 2026-05-13

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeffrey Mensch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a person make a series of choices. Sometimes they pick option A, sometimes option B. You know exactly how much they like each option in different situations (their "payoffs"), and you know the general odds of those situations happening (the "prior").

The big question this paper asks is: Is this person just making choices based on what they know, or is someone else secretly feeding them information to nudge them toward a specific outcome?

In the world of economics, this "someone else" is called a Sender, and the person making the choice is the Receiver. The Sender wants the Receiver to pick a specific action that benefits the Sender, so the Sender designs a clever way to share information (like a news headline, a product review, or a lawyer's argument) to change the Receiver's mind. This is called Bayesian Persuasion.

The problem for researchers is that they often can't see the Sender. They only see the Receiver's choices. This paper provides a "detective's toolkit" to figure out if a Sender is actually at work, even if they are invisible.

Here is how the paper solves this mystery, explained through simple analogies:

The Core Mystery: The "Hidden Puppeteer"

Imagine a magician (the Sender) and a volunteer (the Receiver). The volunteer has to pick a card. The magician knows the volunteer's favorite card, but the volunteer doesn't know which card is in the deck yet. The magician can whisper hints to the volunteer.

If the volunteer picks a card that seems random, is it just luck? Or did the magician whisper a hint that made that card the best choice at that specific moment?

The author, Jeffrey Mensch, says: "We can't just guess. We need a mathematical test to prove if a hidden magician is pulling the strings."

The Two Rules of the Detective Game

To prove a Sender is persuading the Receiver, the data must pass two specific tests (called "axioms" in the paper).

Test 1: The "No Regret" Rule (NIAS)

The Concept: If the Receiver picks a card after hearing a hint, that card must be the best possible choice given that specific hint.
The Analogy: Imagine you are at a buffet. You see a sign that says, "The soup is spicy." Based on that sign, you choose the soup. If you actually hate spicy food, you wouldn't have picked it. If you did pick it, the sign must have convinced you the soup was actually mild, or that you were hungry enough to risk it.
The Paper's Claim: If the data shows the Receiver picking an option that would be a terrible choice given the information they supposedly received, then the whole story falls apart. The Receiver must be acting rationally based on the information they have.

Test 2: The "No Free Lunch" Rule (NBPS)

This is the paper's big new idea. It's the most important part.

The Concept: If a Sender is doing their job, they are only sending information if it helps them get a better result than doing nothing. If they do nothing, the Receiver just goes with the "average" guess (the prior).
The Analogy: Imagine a car salesman (Sender) trying to sell you a car (Receiver).
- If the salesman gives you no brochures, you buy based on your general knowledge (the "prior").
- If the salesman gives you a brochure, it's because they think showing you specific details will get you to buy their car, which is better for them than you just guessing.
- The Test: The paper asks: "Could the salesman have achieved the exact same sales results by giving you a brochure that included a 'blank page' (representing no new information)?"
- If the answer is YES, then the salesman is lying. Why would they bother giving you a brochure if they could have just given you a blank page and got the same result? A rational salesman only gives a brochure if it strictly improves their chances. If you can swap their complex brochure for a "do nothing" option and get the same outcome, the brochure wasn't actually necessary. Therefore, the data doesn't fit the story of a strategic Sender.

The "Menu" Analogy

The paper looks at many different "menus" of choices (different situations the Receiver faces).

The Trick: The author shows that you can't just look at one menu. You have to look at how the Receiver behaves across all menus.
The Metaphor: Imagine a chess player. You can't judge if they are playing perfectly by looking at one move. You have to look at their whole game. The paper says: "If we can rearrange the moves across different games in a way that keeps the final score the same, but somehow includes a 'pass' (no information) where there wasn't one before, then the player wasn't actually playing optimally."

What About "Transparent Motives"?

Sometimes, the Sender's goal is obvious. For example, a prosecutor wants a conviction regardless of whether the defendant is actually guilty or innocent; they just want the result of a conviction.

The paper says: If the Sender's goal doesn't change based on the specific details of the situation (state-independent), the test gets even stricter.
The Analogy: If a politician wants to get elected, they don't care which specific policy you support, they just care that you vote for them. If their strategy looks like it could be replaced by a "random guess" strategy that gets them the same votes, it's not a smart strategy. The paper provides a specific test for this "transparent" scenario.

The "Posterior Mean" Twist

The paper also looks at a special case where the Receiver only cares about the average of the information (like the average price of a stock, rather than every single fluctuation).

The Analogy: Instead of looking at every single temperature reading of a day, you only care about the "average temperature."
The paper shows that even in this simplified world, the same logic applies: If the Sender is doing their job, they can't be achieving their goal by just giving you the "average" guess. They must be giving you something more specific to make their job easier.

Why This Matters (According to the Paper)

The paper doesn't claim to solve real-world problems like "fixing healthcare" or "stopping fake news" directly. Instead, it provides a mathematical tool.

It says: "If you have a dataset where you know what people chose, what the options were, and what the odds were, you can now run a specific calculation (using linear programming) to say: 'Yes, this data is consistent with a hidden persuader,' or 'No, this data is impossible under the theory of persuasion.'"

It turns a philosophical question ("Is someone manipulating us?") into a solvable math problem.

Summary

The Setup: We see a person making choices. We know their preferences. We suspect a hidden "Sender" is feeding them info.
The Test:
- Rule 1: Did they pick the best option for the info they got? (If not, the theory fails).
- Rule 2: Could the Sender have achieved the same result by not sending any info? (If yes, the Sender isn't actually persuading; the theory fails).
The Result: If the data passes both rules, we can say, "Yes, it is possible that a hidden Sender is optimally persuading this person." If it fails, we know the story of the hidden Sender doesn't fit the facts.

The paper is essentially a "lie detector test" for information design, ensuring that if we claim someone is being manipulated by a hidden hand, the math actually backs it up.

Technical Summary: Revealed Bayesian Persuasion

Problem Statement
The paper addresses the empirical challenge of testing whether a Decision Maker's (DM) stochastic choices, observed across various menus and conditional on the state of the world, are consistent with the hypothesis that the DM is being influenced by a "sender" via Bayesian persuasion. Unlike standard revealed preference tests where the DM's preferences are the sole focus, this framework assumes the analyst knows the DM's state-dependent preferences ( $v$ ) and the prior ( $p_0$ ), but does not observe the sender's preferences ( $u$ ) or the sender's existence directly. The core question is whether the observed data can be rationalized as the outcome of an optimal information design problem where a sender ex-ante optimizes their expected payoff by designing a signal structure, subject to the constraint that the DM acts optimally given their posterior beliefs.

A critical non-triviality condition is imposed: the sender is assumed to use a non-trivial persuasion strategy (i.e., providing information distinct from the prior) if and only if doing so strictly increases their expected payoff compared to revealing no information. This rules out degenerate cases where the sender is indifferent between all actions.

Methodology and Framework
The author adopts a revealed preference approach, analogous to Afriat's (1967) characterization of utility maximization. The dataset consists of State-Dependent Stochastic Choice (SDSC) data: a collection of pairs $\{(A, \sigma_A)\}$ , where $A$ is a menu of actions and $\sigma_A$ is the conditional distribution of choices given the state $\omega$ .

The methodology proceeds in three steps:

Revealed Posteriors: The analyst infers the "revealed posteriors" ( $p_{A,a}$ ) associated with each chosen action $a$ in menu $A$ using Bayes' rule and the observed choice probabilities.
Optimality Conditions: The paper utilizes the concavification result of Kamenica and Gentzkow (2011), which characterizes optimal persuasion strategies as distributions of posteriors that lie on the concave closure of the sender's indirect utility function.
Finite Characterization: Drawing on Lipnowski and Mathevet (2017), the author restricts the set of posteriors that need to be considered for optimality to the "outer points" (extreme points of the convex regions where specific actions are optimal for the DM). This reduces the infinite-dimensional problem of checking all possible signal structures to a finite set of potential deviations.

Key Axioms and Results
The main result, Theorem 1, establishes that a dataset is consistent with non-trivial Bayesian persuasion if and only if it satisfies two axioms:

No Improving Action Switches (NIAS): A standard axiom from the information-based choice literature (Caplin and Martin, 2015). It requires that for every action $a$ chosen with positive probability in menu $A$ , $a$ must be optimal for the DM given the revealed posterior $p_{A,a}$ . Formally, the DM cannot switch to another action $b$ in the same menu and achieve a higher expected payoff at that specific posterior.
No Balanced Prior/Outer-Point Signals (NBPS): This is the novel contribution of the paper. It tests whether the observed distribution of posteriors is optimal for some sender preference. The axiom states that there should be no alternative sequence of "Prior/Outer-Point (POP)" signals that:
- Preserves the average state-dependent probability of choosing each action across the menus (maintaining the same joint distribution of actions and states).
- Is Bayes-plausible for each menu.
- Places positive probability on the prior $p_0$ for at least one menu where the original data did not.
The intuition is that if the sender benefits from persuasion, the prior must be suboptimal. If one can construct an alternative signal structure that yields the same average payoffs (and thus the same sender utility for any preference) but includes the prior in its support, the original signal could not have been optimal. The proof utilizes Farkas' Lemma to show that the non-existence of such a "balanced" signal is necessary and sufficient for the existence of a sender utility function that rationalizes the data.

Extensions and Special Cases

Transparent Motives (State-Independent Preferences): In Theorem 2, the paper adapts the test for senders whose preferences depend only on the probability of actions taken, not the state. The condition is strengthened to SI-NBPS, which requires preserving the unconditional probabilities of actions rather than the state-dependent ones.
Posterior-Mean Bayesian Persuasion: In Section 7, the framework is extended to environments where preferences depend only on the posterior mean (Dworczak and Martini, 2019). Theorem 3 provides necessary and sufficient conditions (NIAS and NBPS-M) for this setting, accounting for the fact that optimality is characterized by a convex price function rather than a hyperplane.
Varying Priors: Theorem 4 extends the results to datasets where the prior changes across menus, requiring a uniform version of the axioms (UNIAS and UNBPS).

Significance and Claims
The paper claims to provide the first revealed preference characterization of Bayesian persuasion. Its primary significance lies in offering a tool for empirical researchers to test for the presence of information design in data without imposing parametric assumptions on the sender's preferences.

Empirical Utility: The conditions are formulated as a system of linear inequalities, allowing for computational testing via linear programming. This enables applied researchers to falsify the hypothesis of Bayesian persuasion in real-world datasets.
Theoretical Contribution: By characterizing the set of sender preferences consistent with the data, the paper bridges the gap between theoretical models of information design and empirical observation. It clarifies the distinction between Bayesian persuasion and Rational Inattention (RI), demonstrating that the two theories yield distinct testable implications (e.g., Example 5 shows a dataset rationalizable by persuasion but not by RI).
Robustness: The framework handles incomplete choice data and does not require explicit tie-breaking rules for the receiver, as the non-existence of balanced POP signals implicitly handles indifference.

The author notes that while the theoretical conditions are necessary and sufficient, practical application faces challenges regarding data availability (specifically the need for state-dependent choice data and known priors) and the potential for noise in decision-making, which would require adaptations to the axioms.