Mediated Cheap Talk Design

Itai Arieli, Ivan Geffner, Moshe Tennenholtz

Introduction

In the field of information design, researchers study how an informed "sender" can strategically reveal information to influence an uninformed "receiver." This interaction typically falls into two categories: Cheap Talk, where the sender lacks commitment power and often fails to convey useful information, and Bayesian Persuasion, where the sender can commit to a revelation rule, maximizing their influence.

This paper explores a middle ground: Mediated Cheap Talk. By introducing a trusted mediator and a second informed sender, the authors demonstrate that it is possible to achieve outcomes nearly as effective as Bayesian persuasion without requiring the senders to have any commitment power.

The Power of Two Senders

The core innovation lies in the use of two senders. While a single sender without commitment often leads to "babbling" (where no information is exchanged), a mediator can "cross-check" reports from two senders. If the senders disagree, the mediator can trigger a punishment action, incentivizing both to report the state truthfully.

One of the most surprising findings is prior independence: the set of implementable policies depends only on the players' preferences in various states, not on the initial probability distribution (the "prior") of those states.

Characterizing Implementability and Resistance

To determine which policies can be sustained in equilibrium, the authors define a binary relation over states. They introduce a concept called "resistance" $r(\omega)$, which measures the ratio of the receiver's utility change versus the sender's utility change in a given state.

Visualization of state relations and implementability conditions used to define the mediator's strategy.

By sorting states based on their resistance, the authors developed an efficient $O(n \log n)$ algorithm to identify the optimal equilibrium for the senders. This moves the problem from abstract existence proofs to a concrete, computationally efficient framework.

Results: Bridging the Utility Gap

The effectiveness of this model is demonstrated through a synthetic 6-state example. In this scenario, the utilities for the receiver ($u$) and the sender ($v$) are defined across states $\omega_1$ through $\omega_6$.

Utility values for a 6-state synthetic example used to compare communication models.

In this specific case study, the utility gains are significant:

  • Single-sender Cheap Talk: 0.5 utility.
  • Bayesian Persuasion (Full Commitment): ~0.758 utility.
  • Mediated Two-Sender Model: ~0.754 utility.

This result demonstrates that the mediated two-sender mechanism captures nearly the entire benefit of commitment without actually requiring it, effectively bridging the gap between cheap talk and full persuasion.

Conclusion and Future Directions

The paper provides a robust foundation for mediated communication, though it notes several areas for future research. The current model assumes binary actions and that both senders have full information. Future work could explore "collusion-proof" mechanisms to prevent senders from coordinating on a lie, as well as scenarios where senders possess only partial or overlapping information.

Summary updated

Abstract

We study an information design problem with two informed senders and a receiver in which, in contrast to traditional Bayesian persuasion settings, senders do not have commitment power. In our setting, a trusted mediator/platform gathers data from the senders and recommends the receiver which action to play. We characterize the set of feasible action distributions that can be obtained in equilibrium, and provide an O(n log n) algorithm (where n is the number of states) that computes the optimal equilibrium for the senders. Additionally, we show that the optimal equilibrium for the receiver can be obtained by a simple revelation mechanism.
BibTeX
@inproceedings{arieli2023mediated,
  author = {Itai Arieli and Ivan Geffner and Moshe Tennenholtz},
  title = {{Mediated Cheap Talk Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i5.25678},
}
Figure