Dear YouTube AI, I Am Not a Nazi
I would say about 35% of my feed correctly anticipates my nerd interests in content in mathematical foundations, theoretical physics, and computer science. About 35% of my feed anticipates my interest in various martial arts and combat sports. The rest of my feed is for random walk curiosity topics I’ve dipped into over some recent time period.
There are a strangely high number of gun videos. All sorts of things from guys running around promoting tactical training to gun reviews etc... I like martial arts and combat sports. I don’t like guns. Maybe you like guns. Good for you. We all have our things. Not for me though. Maybe that makes me such a statistical outlier that it just breaks any attempts to mathematical model me. If only there was some way to tell the algorithm it has something wrong. Oh wait – there is. I click “Not Interested” and “Don’t Recommend This Channel” to gun video I see.
As far as I can tell those buttons aren’t connected to anything on the back-end. They seem to exist to provide the user with some comforting illusion of control but have no. This is pretty absurd on its face. I mean we can talk about artificial intelligence until we are blue in the face but when the actual human keeps slamming a button saying “NO THE ALGORITHM HAS THE WRONG OUTPUT TRY AGAIN” and it has zero impact then what kind of choices are being made at Youtube. Like what are all those people actually doing all day. But OK, I’m a big boy. I don’t like guns but I’m not triggered and I gloss on by.
But yesterday I got a whole new level of WTF. Yesterday I got a full on Nazi propaganda video. Not Nazi in the sense the term gets over-used as internet rhetoric. No, I’m talking about the kind of guys who would be proud to be called Nazis. Some doofus called the Leather Apron Club appeared on my feed with a video explaining how overrepresented Jews are on Joe Rogan and how dangerous this is for “us.” It turns out this is a whole lovely channel with 30k subscribers devoted to warning good white Christian American folks about the insidious Jewish media domination conspiracy.
I’m not calling for this to be banned. This isn’t a free speech thing. This is a serious questioning of what the hell any of the Youtube engineers and managers have been doing for the past 17 years. 17 years. Youtube isn’t a new product which popped up last month. They have had almost two decades and virtually unlimited resources to file down the rough edge cases. And yet here I am – a sucker actually sending $12 to Google every month– being recommended Nazi propaganda. Sure I can click “DON’T RECOMMEND” and “NOT INTERESTED” until I’m blue in the face but based on prior experience I doubt that will have any effect. If anything, because I was so puzzled by what was happening that I looked at who this Leather fellow was, Youtube’s trillion dollar “AI” tech stack will probably serve more and more of this stuff. In which case I just exit. I leave Youtube and take my $144 of bloated 99% margin revenue with me. A+ business job managers.
I’m just baffled here. Everyone is terrified of AI taking over the world and here the world’s biggest AI entity after two decades and billions upon billions of spending seems to have a recommendation algorithm with precision I could hack together in Microsoft Excel over a weekend. You have legions of very highly paid business managers who haven’t realized that this is a great way to not make revenue. How is this product such a flaming pile of dog poo that this can systemically happen? This is Hacker News. I’d sure love it if some people with current or former insights from the Youtube team can help explain.
53 comments
[ 1.9 ms ] story [ 128 ms ] threadI use uBlock Origin and haven't seen an ad on Youtube in years.
Here's my theory on why that happens: Youtube knows I'm male, kind of old, a tech nerd, and that I likely don't have any kids or a wife. That's it. It's the kind of content I'm supposed to like in my gender/age/status cohort.
Yes. This is exactly how personalized advertising works.
Advertisers pay to target specific groups/cohorts in a statistical manner. Sorry that you don't like your cohort but this is a well known and unavoidable byproduct of their statistical approach.
It's like when a friend has cancer. You search for it on Google and the next thing you know, you're bombarded with ads for quack remedys.
The only effective solution --- avoid those who play this game.
Advertising doesn't have to work this way. There are alternatives --- such as basing it strictly on expressed interest. If you search for cancer, you should expect some ads related to cancer. What you shouldn't expect is these ads tracking you all over the internet.
Advertising works but there is not much independent evidence to show that personalized advertising is really any more cost effective.
The question for advertisers --- is this really more cost effective than a simple, non-personalized, context based approach?
There is evidence that advertisers are increasingly deciding it is not. For example --- Amazon's context based ads are one of it's fastest growing areas.
https://www.businessinsider.com/inside-amazons-growing-ad-bu...
While YT staff may not actually hold these extremist views themselves, my guess would be they're nevertheless actively pushing it for commercial reasons. Just like it turned out that the supposedly liberal pre-Elon old guard of Twitter had special protections in place that made accounts such as LibsOfTikTok immune to getting banned, I am quite certain Youtube has the same protections for extremist channels that would otherwise run afoul of its T&C on the daily.
The slickest and most subversive form of advertising is that which isn't immediately recognized as such.
https://news.ycombinator.com/newsguidelines.html
― Jean-Paul Sartre
―Joseph Stalin, "Anti-Semitism: Reply to an inquiry of the Jewish News Agency in the United States" (12 January 1931)
https://en.wikiquote.org/wiki/Joseph_Stalin
This isn't a zero sum game: every time you tell it "don't recommend" and "not interested", that is an interaction, too. You are on Youtube instead of some other website or service. This is the attention economy equivalent of a loss leader.
If you believe Youtube no longer serves you, quit. This goes for any other service, too.
Implying YouTube literally ever made money lmfao
On Youtube I can watch IAS lectures on cutting edge topics on quiet evenings. I can find incredible creators explaining mathematical and physics concepts - often with innovative and useful graphical presentation. I can watch Kendo matches from Japan and more BJJ instructional videos from true masters than I'd have time to watch in 10 lifetimes. For someone who spends all day reading/writing text for work and then still reads a lot on free time - well sometimes you just want to watch/listen vs. more text input!
But yes I will move on if anything like this happens again. Nothing is irreplaceable.
> If anything, because I was so puzzled by what was happening that I looked at who this Leather fellow was, Youtube’s trillion dollar “AI” tech stack will probably serve more and more of this stuff.
So... it worked. You seem to be confused about the goals of Youtube's algorithm - it isn't to show you the content you're most interested in, per se, it's to show you the content you're more likely to add value to by engagement. That can and often does correlate to your interests, but controversy works just as well, which is why these platforms often incentivize controversy over quality. Sometimes they'll throw random stuff into your feed just so see if you'll bite.
Consider deleting your view history if you haven't already. Go through your subscriptions and see who they're following. Absolutely do not click on more than one of someone's video if you don't want to get flooded with their content.
- Exactly ... which is both hard and sad. You don't always know what the content is before you click out of curiosity or otherwise. For example, I couldn't write this post before clicking on the Leather Apron guy to see what the heck was going on.
And we should be able to listen to normalize arguments from the "other side" of politics without starting a cascade ending with Nazi propaganda. And the easiest way to do that is to take the user "STOP IT" feedback seriously. Which is really the point of my post if my rant and need to keep things under 4k characters obscured the TLDR.
It is a disservice to the end-user and to all the amazing creators for whom Youtube is as present an irreplaceable venue for distribution.
Purely by its own machine learning, the YouTube algorithm becomes sophisticated at inching people toward the high-watchtime material. There is no mal intent. There's no intent at all except for revenue. The AI doesn't "know" what the content of the reactionary videos even is. All it "knows" is that people who get there watch longer, and that people get there by way of other interests, such as martial arts, history, or guns. It also "knows" that this transition has to happen gradually to succeed. The recommendation progression is subtle:
Martial arts -> combat -> guns/militaria -> liberals are coming to seize your guns
Martial arts -> combat -> guns/militaria -> look at this neat Nazi gun -> were the Nazis so wrong???
Martial arts -> MMA -> Joe Rogan -> Jordan Peterson -> feminism is destroying Western civilization
Martial arts -> MMA -> Joe Rogan -> (((George Soros))) is a space reptile
Your line of questioning assumes that this process is malfunctioning, when in fact it is working as intended. You're correct that it's ethically grotesque, but since when have companies ever taken responsibility for externalities without being compelled to via regulation? YouTube has the legal right to host and broadcast whatever (non-obscene, non-copyrighted) content makes it the most money. The only free-market answer would be advertisers pulling money, which they occasionally do as a disciplinary mechanism. But if the advertisers don't really care, well...
But when a user explicitly says "NO NO STOP YOU HAVE IT WRONG" and it has no impact then it is not optimal design, it is bad design, and not possibly revenue maximizing. Fixing it would also be a great way for these companies to stop stepping in public relations unforced errors.
Agreed completely. Uh… is it your thought that they do not employ left-wing polemics for precisely the same reason?
Because https://news.ycombinator.com/item?id=34463025
The real concern is if Google really has accepted money from this channel in exchange for access to the feeds of people who are into violent content. To say that scenario is believable is an understatement, but to believe it is dark. I'd prefer not to.
I do suspect that the recommendation algorithm for Premium is slightly different and perhaps more favorable to gun content. Pretty much all gun content is demonetized, and anything that’s demonetized gets recommended less; I suspect Premium is an exception to this. Also, there may be more gun content than usual lately because SHOT Show, a trade show for the gun industry, is happening this week.
ML isn't perfect.
That's it. You found a false negative.
But there's an endless amount of new content, and new content has a period before its properly labeled and sorted into the right filter bubbles, so you will likely never get to a world where you don't have to see that stuff.
I mean how else did you know what this leather apron club was? You seem to of described the content of the video that was recommended as if you watched it or clicked on it or interacted with it in some positive manner.
You sent a conflicting signal to YouTube by investigating the content that you don't want to be recommended by interacting with it a whole bunch.
Next time you get recommended things that you don't want to be recommended I'm going to recommend to you to not interact with it at all. Don't click on it don't do anything that is considered engagement to it.
I also use YouTube a bunch and I also pay for YouTube premium. All of my recommends are extremely accurate and that is because the only thing that I view or consume are things that I want to view or consume.
I don't click on things I disagree with or would be offended by I don't click on shorts because I hate that entire concept of content and I don't click on trash content like haha funny viral videos because I am under the complete understanding that interaction and engagement are what drive the algorithm that recommends me content.
I could get totally rabbit holed and click on this stuff that every once in awhile it tries to recommend me that is slightly outside of my bubble in order to research it and find out why it tried to recommend me this. but you know what that would do? It sends a signal to them that I clicked on the video and then I clicked on their homepage to find out more and then I clicked on their about to look at other channels associated with it maybe I clicked on a couple more videos to find out if all of their content is like this maybe I clicked on videos on the side that are related to the video that it thinks I would be interested in if I like the video I'm currently clicked on and all of these are positive signals that feed the recommendation engine to give me more things that I don't want to be given.
And I'll go as far as to say I wouldn't even do that in an incognito tab or from the same IP address or from the same computer because it will somehow leak into my logged in normal account.
I guess I also want to put forth the idea that the negative signal that you put out by saying you don't want to be recommended something is possibly exponentially weaker to the recommendation engine then any amount of investigatory engagement that you do. You say AI a lot and frankly this is how YouTube's recommendation engine has worked for a very long time and I think it's going to be an even longer time and it's going to require more invasive analytics for YouTube to be able to have the ability to pick up on intent of a click. Because that's what this really revolves around. If a human was behind the recommendation engine and you were able to small talk with said human during your frustration of investigating why you're being recommended something then that human would understand to ignore this engagement.
I've been using pockettube addon and just add my favorite channels to groups. Now I can just watch tech, and get my favorite tech shows. Cars, Music, etc. Plus I get uncensored and time/popular feeds sort views again for each of my groups. This is 10000% better than clicking "subscriptions".
Sad people are just herded like sheep on what content they consume.
https://pockettube.io/
Youtube doesn't decide what I see, I decide that. Youtube merely offers me data that I filter.
Seems taking the web into my own hands is the only solution for this, I'm sure Youtube and other sites couldn't care less about their users.
I think I've spoke about this here previously, but I do feel YouTube radicalised me in my early 20s. In my case though it was from the left, not the right, and ack then there was no real recommendation algorithm. There was trending videos and subscriptions, but if I remember correctly the only recommended stuff was basically just related content next to videos.
Personally I don't know if the current recommendation algorithm is any worse than the subscription feed. I think we humans might actually just be drawn to more extreme or "pure" versions of our own beliefs. At least I feel this is true for myself as it's something I have to actively fight all the time.
A good example of people being drawn to more extreme versions of their own beliefs is TV news. People who watch CNN don't watch Fox News and people who watch Fox News don't watch CNN. This self-selection of media content serves to amplify political division and radicalise people because most people will only watch and read what they already agree with when they have to pick the content.
So I think there's an argument to be made here that while not perfect the YouTube recommendation algorithm might actually be better than the old way of consuming content by personal selection. And I guess the fact you were recommended something you so strongly disagree with that you wrote this post you're kind of proving what I've been thinking. This recommendation clearly didn't serve to radicalise you, but for better or worse it might have helped you understand what people you wouldn't normally interactive with actually think on some subject matter. And sometimes you had the perspective, but I find I often that's not the case.
So yeah, while YouTube sometimes recommends me crazy stuff, I no longer have a list of 50 far-left videos waiting for me when I open the site. And while I find what I'm recommended is still biased to my personal preferences, it's less biased then myself.
My personal answer to all of this is that YouTube is a failed product for me. It may serve some general consumption/ad revenue production use case but it is a terrible product for a self-directed “power user.” I see the situation analogous to Twitter. The benefits of the platform are too powerful to entirely leave behind but horribly deficient to what they could be and there is no real alternative because of monopoly. YouTube in my view long ago stopped focusing on providing more value in a two sided market between content consumers/creators and shifted to pure monopoly resource extraction. I’m sure there are alternative emerging platforms but the logic of natural monopolies in two sided markets makes it brutally hard for any of them to reach critical useful scale.
But I realize I’ve gotten very lazy over the past decades and outsourced far too much power to the platform companies. Business is always some combination of providing value to your customers and extracting value for yourself but too many of these platforms have lost the balance. I’m tired of being the product even when I try to provide an alternative paid revenue stream to opt out of that deal.
Annoying and painful as it will be (I’m at a point in life where I no longer have bandwidth to self-hack every thing) I am going to have to reclaim big chunks of my digital autonomy. Thanks all for the great discussion here.