Token-based models often fail to recognize intentionally obfuscated hate speech, such as the phrase "$ ki11 yrslef a$$hole$". Because standard word-level tokenizers cannot process these variations, character-level n-grams are necessary features to detect disguised offensive language.