I would think having humans more involved in training the algorithm could scale much better.
Also, detecting videos that are inappropriate for children is a lot harder than determining certain content creators that are trustworthy to post videos that are appropriate (and to tag them correctly). That can be learned from the user's history, how many times their stuff has been flagged, getting upvotes from users that are themselves deemed credible, and so on. The more layers of indirection, the better, a la PageRank.
So even without analyzing the video itself, it would have a much smaller set of videos it can recommend from, but still potentially millions of videos. You still need some level of staff to train the algorithm, but you don't have to have paid staff look at every single video to have a good set of videos it can recommend. The staff might spend most of their time looking at videos that are anomalous, such as they were posted by a user the algorithm trusted but then flagged by a user that the algorithm considered credible. Then they would tag that video with some rich information that will help the algorithm in the future, beyond just removing that video or reducing the trust of the poster or the credibility of the flagger.
The algorithm works really damn well for 99.999% of the cases. It manages to show me great recommendations from very niche things I'm interested in. But it's the very same behavior that can, in some cases, lead to issues.
are you sure that it's not you who knows very well how to curate their own content and who to subscribe to rather than the recommendation system?
I'm not sure heavy automation is needed here, people jump from content creator to content creator by word of mouth. In contrast most algorithmic suggestions to me seem highly biased towards what is popular in general. I click on one wrong video in a news article and for the next two days my recommendations are pop music, Jimmy Kimmel, Ben Shapiro and animal videos
Not for me, for example I've been watching a few PyCon and I/O talks, and it's been showing me other interesting PyCon talks that are highly ranked. It's also giving me good AutoChess and OneHourOneLife Let'sPlays, both of which I've been very interested in lately.
All three things I just mentioned are fairly niche, comparatively, yet it knows that I've been watching a lot of them lately and is giving me more of it.
I'm reminded of how Google images had an issue where dark skinned people sometimes turned up in a search for gorilla. 99.9% of the time, the image recognition algorithm did really well, but here was a case where an error was really offensive. What was (probably) needed was for there to be a human that comes in and, not tag every gorilla image, but simply to give it some extra training around dark skinned humans and gorillas, or otherwise tweak some things specific to that sort of case, so the chance of it happening was reduced to nearly nothing.
There are probably a ton of situations like that in YouTube, where certain kinds of mistakes are hardly noticed (it shows you a video you weren't remotely interested in), but others can be really bad and need special training to avoid (such as where it shows violent or sexual content to someone who likes nursery rhymes and Peppa Pig).
I mean, that's a pretty easy conversation to have, no? "Hey team, when I search for 'gorilla', black people come up instead. Here, look. That's a problem, right? So I'm going to file this JIRA..."
Except what would really happen is people insisting that there's no actual racism involved, because software cannot be racist, rather the purely technical and politically neutral fact that black people do (chromatically speaking) resemble gorillas more than white people, meaning the algorithm is correct to make the association.
And once it was changed, people would complain that it's just another example of Google being run by extremist far-left ideologues who are ruining the meritocratic purity of tech with their social justice agenda.
Also, detecting videos that are inappropriate for children is a lot harder than determining certain content creators that are trustworthy to post videos that are appropriate (and to tag them correctly). That can be learned from the user's history, how many times their stuff has been flagged, getting upvotes from users that are themselves deemed credible, and so on. The more layers of indirection, the better, a la PageRank.
So even without analyzing the video itself, it would have a much smaller set of videos it can recommend from, but still potentially millions of videos. You still need some level of staff to train the algorithm, but you don't have to have paid staff look at every single video to have a good set of videos it can recommend. The staff might spend most of their time looking at videos that are anomalous, such as they were posted by a user the algorithm trusted but then flagged by a user that the algorithm considered credible. Then they would tag that video with some rich information that will help the algorithm in the future, beyond just removing that video or reducing the trust of the poster or the credibility of the flagger.