Online hate speech endangers minoritized groups and has been linked to offline harassment and violence (M¨uller and Schwarz 2020). Yet the mechanisms by which such toxic discourse spreads and evolves remain context-dependent. In this work, we focus on Brazilian Reddit communities to examine these dynamics at scale, focusing on both banned and non-banned subreddits. Leveraging over 20 million posts and comments, we employ two complementary approaches: Word Embedding Association Test (WEAT) to detect implicit semantic biases, and the Pysentimiento transformer model to classify hate speech in Portuguese-language content. Our results reveal that banned subreddits exhibit strong positive associations with hate-related terminology - particularly sexism - indicating a normalization of discriminatory language. In contrast, non-banned communities show broader but equally concerning patterns of prejudice across multiple hate categories, such as racism, body shaming, and ideological hostility. Additionally, we find evidence of a significant propagation effect: submissions containing hate speech substantially increase toxicity in subsequent comment threads. These findings highlight not only the diverse forms of extremism manifesting within Brazil's Reddit ecosystem, but also critical inconsistencies in content moderation practices. Our work contributes to a deeper, culturally contextualized understanding of online radicalization and offers scalable methodologies for monitoring and mitigating toxic discourse on digital platforms.

PDF