You Tube suspend comments on videos with kids age 13 and younger.

YouTube promised to halt comments on videos with young kids.

YouTube is about to reposition how its massive online video service treats clips for children. Following a record $170 million penalty announced, for violating kids’ data privacy, Google’s YouTube pledged to disable comments, notifications and personalized ads on all videos directed at children. And its machine learning will police YouTube’s sprawling catalog to keep kids videos in line, the company said. One problem: YouTube’s machine learning was supposed to be suspending comments on videos featuring young minors already. It hasn’t.

Comment-enabled videos prominently depicting young kids are still easy to find on YouTube. A single YouTube search for one kids-focused subject — “pretend play” — returned more than 100 videos with comments enabled, all prominently featuring infants, preschoolers and other children young enough to still have their baby teeth. “We invest significantly in the teams and technologies that allow us to provide minors and families the best protection possible,” YouTube spokeswoman Ivy Choi said.

YouTube promised to halt comments on videos with young kids.“We’ve suspended comments on hundreds of millions of videos featuring minors in risky situations and implemented a classifier that helps us remove two times the number of violative comments. We continue to disable comments on hundreds of thousands of videos a day and improve our classifiers.”YpuTube said it would disable comments on videos with young kids following an outcry over a ring of softcore pedophilia. Some videos featuring young children included comments with predatory links. Clicking on the links would transport viewers to other moments in YouTube videos with a minor in a sexually suggestive position. And once you fall in that rabbit hole, YouTube’s recommendation algorithm appeared to feed you more of the same.

So YouTube said it would suspend comments on videos featuring minors who were 13 and younger, as well as on videos featuring older minors who could be at risk of attracting predatory behavior. The changes would take place “over the next few months,” YouTube said then. YouTube would make an exception for “a small number of channels that actively moderate their comments and take additional steps to protect children,” the company said at the time. YouTube videos of older children in scant clothing. The pedophilia-ring scandal earlier this year was triggered by a vlogger exposing predatory links after he searched the term “bikini haul.” In response, YouTube said it would suspend comments on videos featuring children aged 14 to 17, too, if the subject had potential for abuse.

YouTube promised to halt comments on videos with young kids.

A search for “teen bikini haul” videos posted in the last month returned one video by a girl identifying herself as 16 years old, modeling different swimsuits. It has 75 comments. Another was by an influencer who discloses her age — 17 — and her birthdate in the video’s description. Her video, showing off more than a dozen two-piece suits, has 309 comments, including one comment asking her to “show your uncensored sweet tushy.”

Google and YouTube’s scale works in its favor, in some respects. Algorithms need data to learn, and YouTube has more video and data about it than anyone else. Machine learning for video, which essentially looks at videos as collections of still frames, also requires a level of computational power that’s more feasible for a company with Google’s resources.  And one of YouTube’s policy changes announced last week could help its machine learning improve. As part of its settlement with the FTC, YouTube will require uploaders to identify videos that are “made for kids,” it said, effectively introducing more labels on its data.  Algorithms need annotations like these to learn, and the more content that’s getting processed, the more the annotations are necessary, according to Arnav Jhala, a computer science professor at North Carolina State University. Algorithms find patterns and correlations between labels and visible features in the frames.

“The more labels they have, the higher correlation they will have, and on unlabeled video, the algorithms will have a higher accuracy,” he said. “But you are dealing with almost an adversary on the other side.”That is, some uploaders have motives to misidentify their videos.