Content Filtering on YouTube: An LLM Approach for Detecting and Scoring Harmful Content

With over 2.5 billion users, YouTube is one of the largest video-sharing platforms. Although filtering tools exist, there is still a need for more personalized and effective systems to detect harmful content for younger audiences. This paper proposes a new framework using Large Language Models (LLMs) to detect and score harmful content in video subtitles automatically. The system provides personalized insights for parents, improving transparency and allowing for quick, customized content evaluation. Our approach integrates LLMs for scalable content scoring on the presence of harmful elements in a video. We show how this framework improves content moderation for children and offers a scalable solution for YouTube. Unlike traditional methods, our approach provides personalized and dynamic filtering based on text, with plans to extend to multimodal analysis to detect harmful visual content.