Abstract
The rapid growth of AI in content moderation raises major challenges in detecting hate speech and introduces new biases. This chapter explores the limits of binary classification models in addressing the complex nature of online hate. Standard machine learning methods often miss cultural context, evolving language, and the fine line between free speech and harmful speech, leading to unfair outcomes for marginalized groups. Through a mixed-method study using SVMs, deep learning, and transformer models on multilingual datasets, we reveal flaws in current systems. Findings show frequent mislabeling of cultural expressions, racial and gender gaps, and declining accuracy as language shifts. Qualitative evidence highlights social harms, including silencing advocacy and reinforcing discrimination. We stress the need for culturally aware, transparent detection and propose hybrid human–AI moderation to protect users while supporting democratic rights.