Community Notes in Brazil: Consensus, Polarization, and Algorithmic Acceptability

Abstract

This article investigates the operation of Community Notes within the Lusophone sphere, based on the public snapshot from April 7, 2026, from which a corpus of 142,448 Portuguese-language notes was extracted. By integrating language filtering with fastText, multilingual embeddings, topic modeling via BERTopic, Named Entity Recognition (NER), and interpretable utility classification through SHAP, the study demonstrates that 82% of Portuguese notes remain in the needs_more_ratings status and that this suspension is systematically unevenly distributed across topics. The analysis supports the thesis that utility in Community Notes is better understood as algorithmic acceptability rather than factual accuracy, and that the system functions as an infrastructure for the algorithmic governance of public truth.