YouTube: Machines saw 6.7 million videos as offensive at the end of last year

Thanks to machine learning, YouTube says it can take millions of YouTube videos that violate its terms offline before anyone has seen them. In the last quarter of 2017, that happened to 6.7 million videos.

Machines rated those 6.7 million videos as spam or “adult content” before a human reviewer did. Of that number, 76 percent was actually removed before a visitor on the video platform could view it.

YouTube introduced machine learning for identifying offensive videos in June 2017. More than half of videos removed for extreme violence now have fewer than 10 views. In early 2017, before the platform deployed machine learning, eight percent of the violent videos YouTube removed had fewer than 10 views, the rest had more views. With this example, YouTube wants to show that it is getting faster at removing videos.

YouTube emphasizes that deploying machine learning means more people will be involved in the review process, as human moderators must flag videos against the platform’s guidelines. Google aims to employ 10,000 people to review infringing content by the end of 2018. Facebook recently announced that it plans to double the number of people for the same task by the end of this year, in addition to security.

In total, YouTube removed eight million videos in the last three months of 2017 for spam or offensive material. That’s a fraction of a percent of the total number of added videos according to the service.