7 ms·
Setting this case aside because it's sloppy reporting based on what another classifier found, this is a hard problem. Youtube gets too many videos and comments
by dehrmann 6y ago
Setting this case aside because it's sloppy reporting based on what another classifier found, this is a hard problem. Youtube gets too many videos and comments for humans to manually review them. Flagging helps, and Google likely uses ML more complex than just text and classifier (who is commenting, is the post trending, patterns in comments, etc.) When things do reach manual review, remember that humans make mistakes, possibly more often than ML, this could be human error.
You actually want the algorithm to be a black box so that people don't know that you can trick it by only commenting between 5pm and 8pm on an IP in the Netherlands from Chrome, and avoid using words like "cactus."
The important part here is a transparent appeal process, not that ML is only pretty good at these problems.
- phone8675309 6y ago> When things do reach manual review I’ve yet to see a case of the hoi poloi get a manual review - you either are a well enough known YouTuber where they have humans dedicated to you, you know someone who works at YouTube and can intercede on your behalf, or you have to whip up a Twitter mob big enough that it hurts YouTube's image or stirs up people with contacts at YouTube. Even then, sometimes that doesn’t work - look at the Terraria dev. YouTube effectively has no manual review for 99.9% of people which is why its automated systems accuracy is so important.
- rmrfrmrf 6y agoThe press are the manual reviewers, and articles like these are merely examples of the escalation process.
- dehrmann 6y agoBe careful not to confuse manual review with an appeal. When I say "manual review," I mean the classifier isn't sure, so a human double checked, or it was sent to a human to validate the classifier. You're right--appeals are hard to get--but we have no idea how much Youtube manually reviews content.
- polishTar 6y agoIt's a genuinely hard problem, but they do have appeal systems that creators can use that go to human reviewer queues for things like video removals and demonetization even though the first pass is done via automation, but don't think humans aren't immune from mistakes either. As of 2018 youtube had 10,000 manual reviewers (and I imagine it's significantly more than that today), but at their scale, that realistically won't go that far.
- awillen 6y agoPeople forget that they get these free services for free, and a big portion of that is because ML tools are able to cheaply do a really hard job of moderating. You hear about things like this because they're so unusual, which is a testament to the amount of time they work correctly without being noticed. Granted, there have been some issues with people getting banned from Google services for what appear to be wrong reasons and then not being able to get any kind of a human response. Those aren't great, but if you ask people whether they want to pay $10/month for YouTube so human moderators can stop this from happening or get it for free and have the .0001% chance that they get banned purely by accident, they'll take the latter. And that's putting aside the fact that there will still be errors with humans doing it, and there will still need to be the same appeal process to resolve those.
- 0_gravitas 6y agoThe service is not free.
- rmrfrmrf 6y agoJust because the user doesn't need to pay up-front doesn't mean the service operates at a loss (YouTube pulled in $15 billion in 2019, according to a quick search). Furthermore, YouTube Premium is a paid offering. IMO Alphabet just uses YouTube as a playground for its ML experiments because it can moreso than because it needs to for viability.
- abnry 6y agoA YouTuber is a business partner. One that YouTube treats poorly.
- ivanhoe 6y agoNot entirely true. First the platform is free, but also is the content that people provide and without it Youtube wouldn't make sense - so it's not a free service, it's a partnership. Second, for many people YouTube is the primary source of income, and I'm sure there would be plenty of content authors who'd be glad to pay a service with human moderators rather than bots - GIVEN THAT they get to keep access to the same traffic that YouTube provides. That's the key thing, you can't just go somewhere else when all the users are on a few popular platforms. If it's not for that, many high-profile authors who get their videos demonetized on stupidities would have left YouTube long time ago - but they can't, big players keep them in a checkmate position due to the form of monopoly on viewers that they have. And that means you can't just say "If you don't like it, leave, it's a free service", because it's not that simple, as leaving in reality means "give up on your business", and most of people just can't afford that. Youtube has them cornered and thus needs to be careful, since it's not a game, it affects people's lives. For many accidental blocking of channel can mean they'll be living in their cars until the Gods of Google show mercy and fix the error.... and if they do it, since sometimes you don't even get a meaningful reply on what happened, just a template message.
- Retric 6y agoActually watching every YouTube video uploaded would take ~100,000 people which is possible, but quite expensive. Actually watching every video that’s flagged for banning might take ~100 people and seem like a reasonable effort. As such arguments about using automated systems are meaningless, they can pre-filter without having final decision making authority.
- dogma1138 6y agoOver 500 hours of video are uploaded each minute so yeah... even 100,000 people might not be enough to just deal With flags...
- Retric 6y ago> 100,000 people might not be enough to just deal With flags... You’re off by orders of magnitude. 500 * 60 * 24 * 365 = 262800000 hours per year / 1.3x playback speed / 2,000 hours of work per person per year = ~101,077 people to watch every uploaded video. Less if people are watching at higher playback speeds or working significantly more than a 40h week. I would personally be surprised if even 1 in 1,000 videos where flagged by ML at which point you’re talking ~100 people. Though that would be a truly terrible job.
- Sebb767 6y ago> 500 * 60 * 24 * 365 = 262800000 hours per year / 1.3x playback speed / 2,000 hours of work per person per year = ~101,077 people to watch every uploaded video. When you have people watch the video at that fast speed and instantly make a decision, you might as well keep the algorithm - the failure rate won't be too different. Additionally, these people need holidays, managers, infrastructure and an actual flag will probably not instantly be clear - you'll need to point out the specific offence and make a case. You might even need to look things up. Then you'll have to handle appeals and discussions, because changing the AI blackbox to an appeal-less human black box wouldn't improve the situation much[0]. Next, you need people who can understand the specific language of that video - most will be english, but what happens when a Nigerian video is flagged? So you need people from that country or at least familiar with it. Plus infrastructure. Overall, you're probably looking at something approaching 400 or 500 people, with this very low flag rate. Assuming conservative 30k$/year [1], this is 15 million USD alone. Doable, yes, but its not the no-brainer you make it out to be. [0] See this case with FB, which was a human error: https://www.nbcnews.com/tech/social-media/facebook-under-fire-censoring-iconic-napalm-girl-photo-n645526 https://www.nbcnews.com/tech/social-media/facebook-under-fir... [1] Which is probably optimistic for reviewers, but we'll have managerial infrastructure to pull that average up.
- Veserv 6y agoYoutube actually could manually review literally every video. [1] In 2019, users only uploaded on average 500 hours of video per minute. This is ~30,000 hours per hour. If you had people manually watching every video at 1x speed for their entire duration, you would need 30,000 reviewers for each 8 hour shift. Even if we assume that a fully-burdened reviewer costs Youtube an astronomical $80,000/yr, that only amounts to $7.2B/yr on manual review if we assume that they have people literally watching every second of every video. In the same year, 2019, Youtube made a revenue of at least $15.15B in ad revenue alone, so the cost of such a form of manual review only amounts to ~50% of Youtube's yearly revenue which is absolutely within the realm of possibility. Now, I am not saying they could do so cost-competitively or profitably, but it is not directly ludicrous on its face to suggest that they could so if they wanted to. Obviously, just watching every video in real-time a single time is not an adequate review process. Things can be missed, misunderstood, or just take more time to process, so it is likely that more review would be necessary for non-obvious cases. But, on the other hand, it is not necessary to review literally every second of every video. If they only reviewed reported videos and reports were required to specify the timestamp, they would probably instantly reduce this total burden by a factor of 10x-100x (assuming they filter out obvious spam reports). This would allow them to allocate more independent reviewers to the actual non-obvious cases resulting in reasonable accuracy with a human appeals process while reducing their overall cost burden significantly lower than the previously stated ~50%. They could then tier their review process based on the amount of views to allowing them to limit how many resources they spend on unpopular and irrelevant videos to focus on more visible and profitable videos while still providing a base level of review. And, all of this is assuming that they do not have any automated mechanism to directly decide on cases that are trivially obvious, thus reducing their burden even lower to only the contentious cases where the automated systems can not create a clear-cut answer. Frankly, I could go on about various other potential mechanisms they could do to reduce costs while increasing efficiency and avoiding bad actors, but what I have already said should already be sufficient to handle the basic problem. All in all, they could probably institute a reasonable human review process that could handle all reasonable cases for <5% of yearly revenue and frankly I would be shocked if it were anywhere near the high end of that estimate. Even without any automated systems I would put my median cost prediction at 1% of yearly revenue and with even basic automated systems probably close to 0.2%. [1] https://www.businessofapps.com/data/youtube-statistics/ https://www.businessofapps.com/data/youtube-statistics/
- mchusma 6y agoI think it's hard to be perfect but easy to be much better: 1) warn content providers before blocking whenever possible. It's rare something is so bad it requires immediate takedown. Allow people to reply to the warnings. 2) next step is to show a warning to content viewers and allow followers and searchers to find, but not autosuggest. 3) have a transparent appeals process at each step.
- pishpash 6y agoIs it more complex than a text classifier though? If it truly took into account who is commenting, and context, etc., it would have noticed very few chess videos were being flagged for what seemed like racist language. It seems more than likely somebody put in a very crude black box to process text without context. (Or should I say, opaque box, so the horde doesn't flag me?)
- abeppu 6y agoI'm not saying that Google's ML isn't doing some very fancy things but: - Ultimately, any supervised technique will depend on some system that can provide ground truth, which in these applications must be from humans. You can't avoid some humans in the loop and still learn a fuzzy human concept like "is hate speech". But if your aren't gathering that ground truth info in a very deliberate way, you can end up with a range of gaps or artifacts. - There's a broad swath of technical approaches which would consider multiple kinds of information, but ensemble them together such that a very strong signal from one can have a determinative impact. It would be easy to have a situation where if the video itself talks about white/black and attack/threat/defend many times, the score from some language model into a top level ensemble model is very high and the other components may not matter. - But a lot of those other components may often be in agreement because of natural structure. A video says black/white a lot, several other videos on the channel do the same, and a bunch of interested users also look at other channels that have very similar feature vectors, suddenly a content level, channel level and community level signal are all in agreement, and your model can believe it found a subcommunity of hate speech. Incorporating many kinds of info into an ML model isn't that hard ... But doing so in a way which is tightly integrated (eg I'm likely to use a specific meaning of a term in my video which matches a use in a video of a related channel linked in my video description, and that use of that word is also related to what's visible in the video at around that time) is pretty hard. Doing so in a way which distinguishes predictions from confidences is harder. All of which is just to say that it's hard to build a really good black box for a complex problem, especially one based around subjective human concepts, and maybe both we and the black boxes would do better by attaching more uncertainty to their predictions. If it's worth the channel owner spending many hours making videos, and viewers spending many many more watching them, I think it's not crazy to ask that if the model predicts that it's likely disallowed content, that a manual reviewer can take a quick look.