Not Just Oversmoothing: Detecting Echo Chamber Effect in Graph Neural Networks

Oversmoothing is a well-known failure mode of Graph Neural Networks (GNNs). However, most existing diagnostics rely on global aggregation measures that fail to capture the heterogeneous dynamics of message passing. Real-world graphs exhibit pronounced community structure, and message passing operates on two timescales, where representations collapse rapidly within and slowly across communities. This creates a critical gap where intra-community representations can already be indistinguishable while inter-community separation persists, which is a failure mode we refer to as Echo Chamber Effect. To quantify this, we propose to use the Echo Chamber Index (ECI), which stratifies pairwise distances by community membership and formally establishes that global energy diminishes while inter-community separation remains less impacted. The estimated ECI for a graph reveals a surprising failure of common oversmoothing remedies, where rather than escaping the echo chamber, these architectures become permanently trapped in it. The consequences depend on label structure, as when communities align with classes, the echo chamber sharpens node classification, and when they do not, the same collapse makes it provably harder. To address this, we propose Community-Aware Split Propagation (CASP), a lightweight model-agnostic plugin that decouples intra and inter-community aggregation and learns their balance from label structure. CASP consistently improves over diverse backbone GNNs across multiple benchmarks spanning homophilic and heterophilic settings. Our source code is available at: https://anonymous.4open.science/r/CASP-3059/.