Misinformation detection is more useful when you know *where* in a video the false claim occurs, not just *whether* it exists—this work enables fine-grained detection at the segment level rather than video level.
This paper tackles video misinformation by identifying exactly where false claims appear within videos. Instead of just labeling entire videos as true or false, researchers transcribed video audio and annotated which specific segments contain misinformation, creating two datasets with 500+ videos. They trained language models to pinpoint these problematic spans, achieving 68% F1 score.