YouTube inserts explicit AI captions into kids' videos

Occurred: February 2022

YouTube's automated video captioning system inserted explicit language into children's videos, according to a team of researchers.

Having sampled around 7,000 videos from 24 children’s YouTube channels, Rochester Institute of Technology and Indian School of Business reseachers discovered that 40 percent displayed words transcribed by YouTube's Automated Speech Transcription (ASR) service in their captions found on a list of 1,300 'taboo' terms. 

Words like “beach” were transcribed as “bitch”, “corn” as “porn”, and “brave” as "rape". The study also found that 1 percent of the videos included highly inappropriate terms, such as "bitch,” “bastard,” or “penis.” Videos posted on kids' channel Ryan’s World saw the phrase “You should also buy corn” rendered in captions as “you should also buy porn,” and “beach towel” transcribed as a “bitch towel.”

The finding prompted concerns that children exposed to explicit language and adult themes that are inappropriate for their age could lead to confusion, distress, and premature exposure to adult concepts.  Many families use the standard version of YouTube, where these algorithmically-generated captions can be seen.

Operator: Alphabet/YouTube
Developer: Alphabet/YouTube
Country: Global
Sector: Media/entertainment/sports/arts
Purpose: Transcribe speech
Technology: Machine learning; Speech recognition
Issue: Accuracy/reliability; Safety
Transparency: Governance; Black box

Research, advocacy 🧮