Rewarding Engagement and Personalization in Popularity-Based Rankings Amplifies Extremism and Polarization

TL;DR

The paper addresses how ranking systems shape exposure and polarization in online news. It develops a discrete dynamic model of user–algorithm coevolution that blends position bias, stance aligned clicking, and engagement driven updates to rankings, and it tests predictions with static ranking parameter estimation, simulations, and dynamic human-in-the-loop experiments. It finds that personalization and engagement rewards amplify extremism and polarization, with personalization driving polarization more strongly and engagement rewards boosting extremism, evidenced by substantial increases in extreme and same stance clicks under dynamic rankings. These results imply that optimizing for engagement and personalization can inadvertently intensify ideological segregation, underscoring the need for transparent design and mitigation strategies on real platforms.

Abstract

Despite extensive research, the mechanisms through which online platforms shape extremism and polarization remain poorly understood. We identify and test a mechanism, grounded in empirical evidence, that explains how ranking algorithms can amplify both phenomena. This mechanism is based on well-documented assumptions: (i) users exhibit position bias and tend to prefer items displayed higher in the ranking, (ii) users prefer like-minded content, (iii) users with more extreme views are more likely to engage actively, and (iv) ranking algorithms are popularity-based, assigning higher positions to items that attract more clicks. Under these conditions, when platforms additionally reward \emph{active} engagement and implement \emph{personalized} rankings, users are inevitably driven toward more extremist and polarized news consumption. We formalize this mechanism in a dynamical model, which we evaluate by means of simulations and interactive experiments with hundreds of human participants, where the rankings are updated dynamically in response to user activity.

Figures (17)

• Figure 1: Overview of our framework.
• Figure 2: Instance of a page of the "Ranked List and Clicking" step. Users can scroll down to see up to 10 articles.
• Figure 3: Results of the Static Ranking Experiment: (a) aggregated user opinion on the topics, (b) Position bias (H1), (c) Stance alignment in clicking (H2), and (d) Active engagement (H3). Error bars indicate standard deviations.
• Figure 4: Simulated values of attention extremism and polarization, for small increases of personalization ($\lambda$ on y axis, in linear scale) and active engagement reward ($\eta$ on x axis, in logarithmic scale). Each cell is calculated as the average of Eq.\ref{['eq:ext']} and Eq.\ref{['eq:pol']} over 1000 independent simulations.
• Figure 5: Simulation results for Consumption Extremism and Polarization for each topic, in the 4 corner cases introduced in sec.\ref{['sec:simulation']} corresponding to different parametrization of the ranking algorithm. Each cross is obtained as the average and standard deviation over 1000 independent simulations.