Detecting YouTube Scam Videos via Multimodal Signals and Policy Reasoning

YouTube has emerged as a dominant platform for both information dissemination and entertainment. However, its vast accessibility has also made it a target for scammers, who frequently upload deceptive or malicious content. Prior research has documented a range of scam types, and detection approaches rely primarily on textual or statistical metadata. Although effective to some extent, these signals are easy to evade and potentially overlook other modalities, such as visual cues. In this study, we present the first systematic investigation of multimodal approaches for YouTube scam detection. Our dataset consolidates established scam categories and augments them with full length video content and policy grounded reasoning annotations. Our experimental evaluation demonstrates that a text-only model using video titles and descriptions (fine-tuned BERT) achieves moderate effectiveness (76.61% F1), with modest improvements when incorporating audio transcripts (77.98% F1). In contrast, visual analysis using a fine-tuned LLaVA-Video model yields stronger results (79.61% F1). Finally, a multimodal framework that integrates titles, descriptions, and video frames achieves the highest performance (80.53% F1). Beyond improving detection accuracy, our multimodal framework produces interpretable reasoning grounded in YouTube content policies, thereby enhancing transparency and supporting potential applications in automated moderation. Moreover, we validate our approach on in-the-wild YouTube data by analyzing 6,374 videos, thereby contributing a valuable resource for future research on scam detection.