Algorithmic Content Recommendations on a Video-Sharing Platform Used by Children
Radesky's team analyzed 2,880 YouTube video thumbnails surfaced across 20 sequential algorithmic recommendations, starting from search terms popular with school-aged children such as Roblox, memes, and well-known creators. Coders flagged thumbnails for features like visual loudness, drama and intrigue, lavish wealth or wish-fulfillment imagery, disturbing or creepy content, violence or pranks, and gender stereotypes. These attention-grabbing features were common from the very first recommendation and stayed roughly as common through later recommendations, while thumbnails using gender-stereotyped imagery became significantly more frequent as the recommendation chain continued. The authors describe this as evidence that creators competing for algorithmic visibility gravitate toward provocative thumbnail imagery, and that the recommendation system does not filter this out for child-initiated searches, associating the platform's incentive structure with children's routine exposure to this kind of content rather than claiming any single video causes harm.
Of 2,880 coded thumbnails, 91.5% used a "drama and intrigue" attention-capturing style, and 79.1% showed general "visual loudness."
Nearly half of thumbnails (44.6%) were coded as "creepy, bizarre, and disturbing," and 40.6% depicted violence, peril, or pranks.
Radesky, Jenny, Enrica Bridgewater, Shira Black, August O'Neil, Yilin Sun, Alexandria Schaller, Heidi M. Weeks, and Scott W. Campbell. "Algorithmic Content Recommendations on a Video-Sharing Platform Used by Children." JAMA Network Open 7, no. 5 (May 2024): e2413855. https://doi.org/10.1001/jamanetworkopen.2024.13855