A machine learning model to automatically identify fast-paced online videos for children

Abstract

Young children now spend most their screen time watching online videos, with platforms like YouTube dominating digital media use among those ages 0–8. Many of these videos are fast-paced—featuring rapid scene changes, frequent motion, and quick-cut editing. Prior research has linked fast-paced media to diminished attention and executive function in early childhood. Yet studies on the effects of fast-paced content focus on television shows or lab-edited clips, failing to reflect the dynamic and diverse nature of online videos. This study introduces a new, scalable approach to identifying fast-paced content in online children’s videos by training a machine learning model on a large, ecologically valid sample of 426 YouTube videos viewed by children under age 3.

Using optical flow, a computer vision technique that quantifies motion across frames, the model accurately classified videos as fast- or slow-paced with 85% precision for fast-paced content and 83% for slow-paced content—substantially outperforming chance. Unlike traditional approaches to coding fast-paced content that focus on discrete edits, our model captures the continuous motion signals that may better reflect how children experience pacing. This work lays essential groundwork for future research on persuasive design features in children’s media. It also opens the door to automated tools that can support families, designers, and researchers in fostering healthier digital environments for young children.