Algorithmic Curation, Filter Bubbles, and Democratic Information Exposure:An Empirical PLS-SEM Investigation with a Multigroup Analysis

Abstract

Background: The proliferation of algorithmic curation on social media platforms has intensified scholarly and policy debates regarding its effects on information diversity and democratic discourse. Despite two decades of theoretical development since Pariser's (2011) seminal filter bubble hypothesis, robust empirical evidence based on validated structural models remains scarce, and most published research relies on computational audits that cannot isolate individual-level psychological mechanisms.

Objective: This study develops and empirically tests a structural model of Perceived Filter Bubble Formation (PFBF) and its downstream effects on Political Polarization Intention (PPI) and Digital News Engagement (DNE), with digital media literacy as a moderating variable.

Method: A cross-sectional online survey was administered to N = 412 adult social media users (Mage = 29.4, SD = 7.8; 54.4% female) recruited through stratified quota sampling across five Latin American countries. Partial Least Squares Structural Equation Modelling (PLS-SEM) was employed using SmartPLS 4.0. Multigroup analysis (MGA) compared platform type (feed-based vs. recommendation-driven) and age cohorts.

Results: The measurement model achieved satisfactory reliability (Cronbach's α ≥ .82; composite reliability ρc ≥ .86) and validity (AVE ≥ .52; HTMT < .85). The structural model explained 61.3% of the variance in Political Polarization Intention (R² = .613) and 44.7% in Digital News Engagement (R² = .447). Perceived algorithmic homogenization (PAH) was the strongest predictor of PFBF (β = .541, p < .001), and PFBF fully mediated the effect of platform usage intensity on PPI (indirect effect = .231, 95% CI [.187, .289]). Digital media literacy significantly moderated the PFBF → PPI path (β = −.198, p = .004), attenuating polarization susceptibility among high-literacy users. MGA revealed significantly stronger filter bubble effects on recommendation-driven platforms (TikTok, YouTube) than feed-based platforms (β difference = .147, p = .012).

Conclusions: Algorithmic filter bubbles operate as a psychologically mediated process conditioned by platform architecture and individual literacy. Platform-differentiated governance interventions and digital media literacy programs constitute evidence-based policy levers for mitigating democratic communication risks.