Tackling Hateful Video Content with Multimodal AI

As online video platforms become central to digital communication, moderating hateful video content has emerged as a critical and complex challenge. Unlike text or image-based content, hateful videos often operate across multiple modalities, visuals, audio, and textual overlays, and carry subtle or culturally specific signals that evade detection. In this talk, I will present two recent studies that aim to advance the frontier of hateful video moderation through multimodal and cross-modal learning. First, I introduce MultiHateClip, a multilingual benchmark dataset comprising 2,000 annotated short videos from YouTube and Bilibili. It provides nuanced labels, hateful, offensive, and normal, along with annotations on target groups and contributing modalities, revealing cultural distinctions in how hate is conveyed across English and Chinese content. Second, I discuss a novel cross-modal transfer learning approach that leverages more easily accessible hateful meme datasets to enhance video moderation systems. This framework includes a human-in-the-loop re-annotation process to align meme and video labels, and demonstrates how vision-language models fine-tuned on re-annotated memes can outperform video-only baselines. Together, these works highlight the promise of data-efficient and culturally aware multimodal AI models for online harm detection and open new pathways for scalable and globally relevant video moderation strategies.