Rational Silence and False Polarization: How Viewpoint Organizations and Recommender Systems Distort the Expression of Public Opinion

TL;DR

The paper addresses how AI-powered recommender systems distort public opinion by coupling individual expression with organizational signaling and platform mediation. It develops a game-theoretic, nested framework that includes fixed private opinions, continuous rhetorical intensity, and two types of viewpoint organizations (participatory and ideological) whose signals alter beliefs and participation. A key finding is that rational silence among moderates can yield false polarization, and recommender-driven community formation can further skew observed opinions toward extreme groups. The work also proposes mitigation strategies—such as tailoring moderation and prioritizing signals from participatory organizations—and discusses practical implications for policy, platform design, and AI training data. Overall, the results highlight a principled route to understanding and potentially correcting distortions in the digital public sphere without requiring changes to underlying private opinions.

Abstract

AI-based social media platforms has already transformed the nature of economic and social interaction. AI enables the massive scale and highly personalized nature of online information sharing that we now take for granted. Extensive attention has been devoted to the polarization that social media platforms appear to facilitate. However, a key implication of the transformation we are experiencing due to these AI-powered platforms has received much less attention: how platforms impact what observers of online discourse come to believe about community views. These observers include policymakers and legislators, who look to social media to gauge the prospects for policy and legislative change, as well as developers of AI models trained on large-scale internet data, whose outputs may similarly reflect a distorted view of public opinion. In this paper, we present a nested game-theoretic model to show how observed online opinion is produced by the interaction of the decisions made by users about whether and with what rhetorical intensity to share their opinions on a platform, the efforts of organizations (such as traditional media and advocacy organizations) that seek to encourage or discourage opinion-sharing online, and the operation of AI-powered recommender systems controlled by social media platforms. We show that signals from ideological organizations encourage an increase in rhetorical intensity, leading to the 'rational silence' of moderate users. This, in turn, creates a polarized impression of where average opinions lie. We also show that this observed polarization can also be amplified by recommender systems that encourage the formation of communities online that end up seeing a skewed sample of opinion. We also identify practical strategies platforms can implement, such as reducing exposure to signals from ideological organizations and a tailored approach to content moderation.

Figures (7)

• Figure 1: Each player with a private opinion $o_{i}$ is matched with another player drawn from an opinion distribution $f$. With probability $n$ they are matched with someone from their in-group and probability $(1-n)$ from out-group. Each player strategically chooses their rhetorical intensity $\gamma_{i}, \gamma_{-i}$. If the player is matched with the in-group, utility is boosted by a factor $\lambda_{\text{in}} > 1$ and, conversely, utility is dampened by a factor $\lambda_{\text{out}} < 1$ when matched with an out-group member.
• Figure 2: The net utility of expressing an opinion (after deducting the cost) as a function of rhetorical intensity for the focal player. The utility-maximizing rhetorical intensity of the focal player decreases with an increase in other players' rhetorical intensity.
• Figure 3: Best response curves of the optimal rhetorical intensity for the focal (shown in red) and non-focal (shown in blue) player as a function of their opinion drawn from an opinion distribution of equal support for approval and disapproval. For the purpose of illustration, the constants $\alpha,\lambda_{\text{in}},\lambda_{\text{out}}$ are set to 0.7, 2, and 0.5, respectively. The intersection of the two curves represents the Bayesian Nash equilibrium rhetorical intensity in the ex-ante form. The plots are shown for different opinions (types) of the focal and non-focal player, and we see that as the opinion moves to the extreme, the symmetric equilibrium rhetorical intensity becomes higher.
• Figure 4: Schematic representation of the organizational stewarding process. (1) An organization randomly samples a signal that conveys information about approval or disapproval group's opinions from the organization's signaling scheme. (2) The focal agent updates their descriptive belief based on the signal (3)). The focal agent estimates the equilibrium rhetorical intensity and expresses opinions. (4)) A community of agents also expresses opinions based on organizational signals. The focal agent and others updates descriptive beliefs based on the expressed opinion in the community.
• Figure 5: Reward heatmap that shows the optimal signaling policy for different viewpoint organizations and for out-group and in-group signals. Brighter regions (yellow) represent higher reward. The $y$ axis shows the organizations' estimate of the true out-group/in-group belief and the $x$ axis shows the corresponding signals to generate. We see that the optimal signaling distribution follows different patterns for both organizations: ideological organizations get higher reward for more extreme signals about the groups' opinions whereas participatory organizations get higher reward from moderate signals.

Theorems & Definitions (1)

• Theorem 1