I was listening to a left-leaning podcast that had on a technology expert who didn’t actually work for YouTube but claimed this was how YouTube monitored for hate speech.
Basically the idea is, all videos are automatically given automatically generated subtitles by computer voice analysis (you can see this by hitting the CC button on the bottom of a YouTube video and clicking “Auto-Generate Subtitles” if the video didn’t have subtitles uploaded already), this also happens to Streamers. Uploaded videos that’s subtitle scripts include “controversial” words like COVID, Insurrection, Boogaloo, and Roseanne. With videos uploaded, they’re flagged for an actual human to watch the content and approve it but in the meantime the video will be automatically let onto YouTube, however if there’s mass reporting on an already non-approved video the video is automatically taken down pending an investigation.
Similarly, streams are monitored for the same words, but since it’s a stream, if a controversial word or words are said in rapid succession the stream will actually be automatically taken down until a human is able to quickly monitor it on the fly and see if the stream should be approved and put back up or permanently taken off the air.
I don’t think so. Given the dozen plus Covid denial videos in the crazy part of my subs and knowing at least one Covid denial debunking video that was blocked multiple times.
One of the channels I watch (fountain pen reviews) has said that mentioning COVID and possibly other things will de-monetize the video, which I assume means that whatever money they might have gotten from it, they won’t get. I have no idea how Youtube videos are monetized except through advertising (which I never see due to Adblock Plus) so I don’t know if that means that ads don’t run, or that the ad revenue is sequestered somehow. Anyway, de-monetizing videos seems to be a pretty effective threat, but I can see how it wouldn’t keep true believers from posting those things.
The main part I would disagree with is having a human watch the video. I see no evidence they do that, as it makes too many mistakes. I suspect they feed the speech-to-text subtitles to an AI which has been trained to try and spot problematic content (by giving it human curated examples). I suspect a human would only get involved if you appeal, or possibly if someone manually reports the video.
There is a similar process with demonetization. But then it would be using the subtitles and AI to try and categorize the content. And advertisers can choose not to put their videos on certain categories of videos. That one is probably set up to err on the side of not running the ads, while the one in the other paragraph is probably designed to err the other direction.
One thing I have heard from other YouTubers is that it’s only really the first few minutes of the video they have to worry about the latter on. Curse words early on can get you demonetized, but later on in the video they’re okay. Though this may just be a quirk of the AI trying to learn. And I have no idea if this would also be true of the top stuff.
Finally, I wouldn’t be surprised if both of these also monitor comments and the description and other signals like that.
Nobody outside of YouTube knows how the algorithms actually work. A lot of the time, when you hear people saying it’s this or that, unless they actually learned that in a published source from YouTube (such as the terms, or the insider channel), they’re guessing.
Sometimes people try to back up their guesses with badly designed experiments where they publish a video or a post with certain things in it, and/or ask their audience to act in certain ways in viewing and commenting on it, but as they don’t have actual control of any of these variables, the results of the experiments may look about right, but perhaps not for any of the reasons asserted.
YouTube has said in the past that the algorithms are looking at keywords in the title and tags. and the mere presence of the word ‘COVID’ made the difference between monetizing a video, or not, in the early days of the pandemic (and unlike some of the other things, this was really easy to empirically test on unpublished content - remove the word from the title, the icon goes green - replace the word, the icon goes yellow)
I don’t think they have stated that the auto-generated captions are a factor - it seems possible, but at the same time, the auto captions are often very badly wrong, so that would probably lead to both false negatives and positives.
Creators are required to rate the suitability of their own content when they upload it - ostensibly for ad suitability, but most of the sensitive topics are covered. Content and commentary can be reported and flagged by anyone with an account - it seems likely that these two things might be used in combination - if you rate your own content as safe, and you get few reports, and those few reports you get are reviewed and dismissed, the system might be more likely to trust your future self-assessments.
There’s a phenomenal amount of data being created on YouTube every minute - for YouTube to monitor things at great depth, would be quite costly in terms of resources - there’s a difference between something being technically possible, and that thing being cost effective to do. Most likely, YouTube only addresses things that are either common enough to erode their profitability and likely to happen en masse, or are considered serious enough individually to bring about another adpocalypse.