In its most recent published quarterly report, YouTube outlines how, between July and September of 2017, it purged the streaming network of over 1.7 million channels, 7.8 million videos, and over a quarter of a billion comments that the company said violated community guidelines outlined in the terms of service. The intriguing part of this though, is that 99.5% of those were initially removed by a censoring algorithm and not by a human decision.
“We’ve always used a mix of human reviewers and technology to address violative content on our platform, and in 2017 we started applying more advanced machine learning technology to flag content for review by our teams,” states the YouTube Blog. “This combination of smart detection technology and highly-trained human reviewers has enabled us to consistently enforce our policies with increasing speed.”
Eighty-one percent of the 7.8 million videos removed were done so by the learning algorithm, and of those, 81% never saw a single play before being rumored. Over three-fourths of the videos were flagged for being spam, while 10 percent was taken down for violating child safety and adult content rules.
And while many videos that promote what YouTube considers to be extremism or violence, the report shows that less than one half of one percent were actually removed for that reason.
But according to the report, what YouTube reports as “spam” also includes videos that they consider “misleading.” That’s pretty vague. The good news, is that YouTube has removed millions of videos that promoted terrorism, or highlighted activities of terrorists, and protecting children with child safety rules will always be a popular move.
When we detect a video that violates our Guidelines, we remove the video and apply a strike to the channel. We terminate entire channels if they are dedicated to posting content prohibited by our Community Guidelines or contain a single egregious violation, like child sexual exploitation.
As for channels, well, that’s where things get more interesting. Almost eighty percent of the channels purged from the system were done so after picking up a third strike, and YouTube also removed channels that they declared were “wholly dedicated” to the single purpose of violating YouTube’s community guidelines. Whatever that means. Eighty percent was taken down for promoting spam, 12 percent for hosting adult content, and just less than five percent for violating child safety rules. As these channels usually had multiple videos in their catalog, YouTube reported that an additional fifty million videos were removed due to channel terminations.
YouTube relies on teams around the world to review flagged videos and remove content that violates our terms of service; restrict videos (e.g., age-restrict content that may not be appropriate for all audiences); or leave the content live when it doesn’t violate our guidelines.
This all sounds great on it’s face, but Google has received push-back from high profile channels, and even members of Congress, that claim that they were targeted, not because they were advocating extremism of violence, but because their channel held views that Google executives didn’t agree with. They claim that members of the “trust flagger program” all share similar views, and as such, only represent one point of view when it comes to judging whether a video or channel has violated YouTube’s community guidelines. Trusted Flagger program members include individuals, Non-Government Organizations (NGOs), and government agencies.
I can understand, that how an algorithm flags a video will evolve over time, because it’s learning to see what is and isn’t a violation. As it learns, then in theory, it’s flagging system will become more accurate. But when it comes to the Human committee that reviews videos and appeals, if there’s an imbalance of opinion, that doesn’t bode well for a free and open internet.
All told, I think YouTube is doing an admiral effort at what could be a nearly impossible job. But who watches the watchers?
You can read the report, including how their guidelines were used, here.