Videos often combine visual data with kinematic or audio cues. If a sensor fails or a file is corrupted, standard AI fails.