Why would someone expect the recommendation algorithm to really know anything regarding whether videos violate their own policies? If the platform already knew the videos violated their policies, presumably they'd already be blocked/removed in the first place. It's not magic.
I think you're missing the point. According to TFA, the YouTube algorithm likes content that breaks their policies very much:
Recommended videos were 40% times more likely to be regretted than videos searched for.
In 43.6% of cases where Mozilla had data about videos a volunteer watched before a regret, the recommendation was completely unrelated to the previous videos that the volunteer watched.
YouTube Regrets tend to perform extremely well on the platform, with reported videos acquiring 70% more views per day than other videos watched by volunteers.
So it's not that the algorithm is just "blind" to content that breaks their policies, it seems to prefer them over other content.
Reading the report, this wasn't actually a random sampling of typical users. They used opt-in reporting from a self-selected group of volunteers who were interested in reporting "regrettable" YouTube recommendations.
If some of these users were deliberately scouring YouTube for objectionable content, it follows that their recommendations would start matching similarly objectionable content for them.
After reading the report, I don't see how any of this is representative of a typical user's YouTube experience. It certainly doesn't resemble my own experience at all.
Isn't it likely that it prefers them because they're more likely to be fully watched/upvoted/commented on/otherwise engaging? Presumably, that fact together with objections to the content are the reason to have a policy...
> Why would someone expect the recommendation algorithm to really know anything regarding whether videos violate their own policies?
There's a big push to blame recommendation engines in the wake of social media misinformation about elections and COVID-19. There's a growing perception, verging on conspiracy theory, that big social media companies are deliberately tuning their recommendation engines to promote this misinformation while publicly claiming to do the opposite.
This Mozilla report feels like a strangely clumsy attack piece. They relied on self-selected volunteers to proactively report videos that they called "regrets" using a Mozilla plugin. Users were given little direction about what to report as a "regret" but according to the Mozilla report they marked as many as 1 in 8 videos as "should not be on YouTube".
I have no idea how these self-reporters were going down a YouTube recommendations rabbit hole that ended up with 1 in 8 recommendations being policy-breaking or inappropriate, but that couldn't be farther from my own experience. I stream YouTube in the background while working on projects around the house and I can't recall the last time I was recommended a strange video that was somehow policy-breaking.
The YouTube algorithm is clearly very self-reinforcing. I suppose if the volunteers were actively searching YouTube for "regret" videos that broke policy then it's likely that the recommendation engine would be triggered into serving up more, similar videos. That's obviously not a valid representation of the typical user experience, though.
> There's a growing perception, verging on conspiracy theory, that big social media companies are deliberately tuning their recommendation engines to promote this misinformation while publicly claiming to do the opposite.
I don't think that this is the perception of many people. I think that many people think (and I think so too) that social networks maximise for engagement quite deliberately. Content that is very engaging can be very good (amazing educational resources) or relatively harmless (cat videos). But, due to human nature, "engaging" videos also include divisive content, propaganda, conspiracy theories, group-think or even just stuff that is highly addictive.
The complaint, therefore, is not that YouTube etc. are actively pushing conspiracy theories—it wouldn't be in their interest to do so necessarily. The issue is that, by trying to maximise engagement, they very deliberately push all the subconscious buttons that may make us behave in irrational ways without caring for the psychological and social cost that this implies, and that any attempt of those networks to curtail the spread of problematic content will never be adequate as long as those underlying mechanisms are still the same.
>I stream YouTube in the background while working on projects around the house and I can't recall the last time I was recommended a strange video that was somehow policy-breaking.
I have a fairly curated stream of videos due to years of Youtube use. I never see objectionable content in my recommended segment. I mostly have videos related to my career area and a few creators who focus on games I play.
However, I've logged in on other computers, or via VPNs on different devices and found that if I just started with a random folksy video (let's say a bandsaw restoration), I would be recommended objectionable (overtly anti-black or anti-Semitic material) very rapidly. What's more, this random walk would then seed my recommended videos with a high proportion of similar videos. I'm not sure how much directed viewing would need to occur to re-adjust the recommendations. This was such a large issue that at certain offices I worked at, I'd pre-seed my youtube history with a few 'chill cafe background music' videos first to prevent a re-occurrence of the time when I walked away from my computer after logging onto a new terminal and opening up what I figured was some decent background noise, then coming back to a pro-hitler historical documentary.
This may be because of my location, demographic, activity on the network or the selection of initial videos. My experience is nothing more than an anecdote, but it leads me to think that extreme content does well with respect to engagement, and accordingly is promoted by an engagement focused process. Is this correct? I don't know. I just know it isn't far fetched.
Say I go over to Fred, and ask him for a book recommendation. I've known Fred for a while, and he's given me good book recommendations in the past, so I usually listen to what he says. Except this time, Fred says that the book I'm most likely to enjoy is Mein Kampf. I'm flabbergasted, because this doesn't sound like the Fred I know. Fred says that he has no idea what the book is about, but there's a small number of people who vigorously recommend it, and they're just so gosh-darned enthusiastic about it that now he's recommending it to anybody with even the slightest amount of interest in politics. He knows that I just read Elie Wiesel's Night, so it seemed right up my alley to Fred.
Swap out "Youtube" for "Fred", and maybe that helps to show the issue. I'm not making an exaggeration here. Youtube has recommended that I watch Nazi propaganda, because I had just recently watched videos debunking those exact propaganda films.
When "recommendations" become "recommendation engines", and the human steps out of the loop, the recommender still has a moral responsibility for the things they recommend. Without a human in the loop, it's a lot easier to close one's eyes to that responsibility. Maybe the solution is to go back to things that don't scale, and keep humans in the loop. Maybe the solution is to spend just as much effort on the quality of the recommendations as the addictiveness of the recommendations. But the current state is not sustainable.
I think your analogy supports the GP's point. The recommendation engine doesn't understand that the recommendation is a violation, it makes a heuristic error because there is an indomitable mountain of content to classify and it lacks the sophistication to accurately do so. It's certainly logical to hold YouTube accountable for the content it recommends, but it's also naive to expect it to achieve the impossible.
> Fred says that he has no idea what the book is about…
And that’s the thing people don’t understand when they blame YouTube. If Fred is YouTube, and Fred doesn’t know that Mein Kampf is propaganda, how is YouTube supposed to remove it?
You’re just reconfirming OP’s point: YouTube doesn’t know that a video violates its policies.
Fred has neglected his duty to know what he is recommending, in the same way that YouTube has neglected their duty. The key part I was trying to get across with the analogy wasn't just that Fred didn't know what he was recommending, but that he didn't do his due diligence before giving a recommendation.
I feel everyone replying to my comment is missing the overall picture.
YouTube has content guidelines... that doesn't mean they are obligated to enforce them, it only means they have easy reasons to moderate content where they feel like it. Their only real interest is profit, generally through showing you ads (by keeping you glued to the screen, sometimes by showing you controversial content) and by harvesting data. I'm not saying that's "right" in a moral or ethical way, it is simply the truth. If you recognize that, then it doesn't matter that they are recommending videos that ostensibly violate their own terms, and arguing otherwise seems to make you just willingly ignorant. YouTube has no incentive to incorporate "regrettability flags" into their algorithm, except to the degree that it serves them.
I'd go further to say that YouTube probably could do a better, more thorough job of cleansing content on their platform, without any additional cost. But having a certain amount of content that is controversial, which panders to a vocal minority, etc actually serves their platform. They use the 1st amendment sometimes to justify this, but they actually don't care one bit about any of that. All they care about is money, and at this scale the dynamics get kinda interesting.
So, your original post was a question, and I answered it.
But maybe your question was rhetorical and masked an overall point you are now trying to clarify. But I still don't understand the point your making!
Are you trying to absolve YouTube of any moral responsibility, while personifying with them with agency ("they have reasons", "they feel like it", "the degree that serves them", "they care") and granting them rights ("1st amendment")?
As for what really motivates corporations - I believe survival, not profit, is the essential goal. Profit is just one strategy - but another is not poisoning it's own environment.
> Youtube has recommended that I watch Nazi propaganda, because I had just recently watched videos debunking those exact propaganda films.
When you're interested in this field and watched documentaries about it, it is not too far off that you want to watch the source material.
Now, it depends a bit on whether we're talking about "modern" Nazi propaganda or 1930's Nazi propaganda - the latter one can be pretty clearly categorized as historical interest, while the former might a bit more problematic. I don't want to say that showing countering view points is generally a bad thing, but in this case it's clearly crossing a line.
It, however, also totally fits in your analogy: If someone is very interested in politics, it makes sense for them to possibly want to read works of dictators and the likes. Not to get radicalized, but to know what happened and be able to prevent history from repeating itself.
I think the point you're trying to make is that the recommendation algorithm should check videos for violations, but then GPs point still stands: Would YT be able to determine the policy violation, the video would probably be already taken down long before it reaches recommendations.
> Now, it depends a bit on whether we're talking about "modern" Nazi propaganda or 1930's Nazi propaganda - the latter one can be pretty clearly categorized as historical interest, while the former might a bit more problematic. I don't want to say that showing countering view points is generally a bad thing, but in this case it's clearly crossing a line.
In this case, it was modern Nazi propaganda. This particular incident was a few years ago, but the title implied that it would be going through the history of a particular dogwhistle. I naively assumed that the history of that dogwhistle would be used as an example of the early steps for de-humanization of Jewish people, how to recognize those early signs of racial hatred, and what can be done to best combat that hatred. Instead, from the outline in the first 30 seconds of the video, the speaker was going through different atrocities committed, and lamenting how limited each atrocity was. It was pure Nazi propaganda, disguised as a lecture, and should not have been recommended to anyone.
A human could have recognized it as such, had a human been within the decision-making loop for recommendations.
> Would YT be able to determine the policy violation, the video would probably be already taken down long before it reaches recommendations.
I think systems tend to proceed based on the incentives and rules that are set up within that system. Youtube's recommendations are designed to increase engagement, regardless of societal cost. I agree that if a human had seen the video, it would have been taken down. But that will rarely be the case, because there are minimal incentives for having good recommendations, as compared to the incentives for having engaging recommendations.
Much of this is due to the push for automating everything, even if that automation results in poor results. I am of the opinion that if something cannot be done correctly at large scale, then it ought to be done only at small scale.
It seems like a reasonable assumption that, as of 2021, the majority of people who read Mein Kampf or watch Triumph des Willens are not Nazis, but rather people who are curious about the history of Nazi Germany for normal, good-faith reasons.
If your personal attitude towards the Nazis is "never again", as it should be, you ought to be interested in how it happened the first time. Part of how it happened the first time is that a lot of Germans actually voted for them and many of the others just sort of went along with the program. Why would they do such a thing? How do I make sure I'm not fooled the same way they were? These are some haunting questions.
I, myself, have watched this sort of primary source material on YouTube, e.g. subtitled videos of some of Hitler's speeches. You might assume that Hitler spent most of his speeches ranting and raving, but he spent a lot of time hiding behind the mask of a responsible politician promising to reduce unemployment. I think watching that stuff helps to answer those questions. And, as I recall, I don't think I actually went out of my way to search for it, either.
Sure, that's plausible, but it seems equally likely that Youtube _can_ detect videos that violate their policy and yet chooses to wait until they get some threshold of reports/media attention because those videos are also very engaging. Having the policy allows them to react when they get pressure without impacting the bottom line.
I wouldn't say equally likely. They both have a non-zero probability of being true but a pattern of malfeasance such as you describe would be detected pretty quickly. Imagine this or thousands of similar conversations: "Brad, have you noticed that our porn videos get removed only when they get over a certain view count?" "I had noticed that, Bob! I wonder if that observation would be interesting to the media?!"
It also has minimal upside because YouTube makes more money on high view count videos by far, so removing only when a certain level of fame has hit a video cuts off the part that is economically interesting to YouTube. The tail is a loss leader.
"Brad, we can't deploy your conspiracy theory classifier because it would reduce the number of high value viewer engagements by 2.7%. Let's just stick with the version that depends on accumulating reports. It's not a big deal because we can always intervene if something embarrassing slips through."
I'm unaware of any evidence consistent with this version of how things work there. By all accounts, YouTube are unlenient with bans and blacklistings. Can you please cite a source?
The original comment I was replying to was speculating about how they imagined youtube worked without evidence. All I did was provide an alternate explanation. I wouldn't even go so far as to suggest it's a bad thing that social networks avoid using more aggressive classifiers, just that I'm skeptical of the "we don't have the ability" explanation.
By all account YouTube are arbitrary with bans and blockings. That's not the same as unlenient.
The standard example of the phenomenon I'm describing is from Twitter. They successfully keep nazi content out of German twitter, but not the rest of the world. That suggests that Twitter's moderation is partially constrained by their own policies, not just technical ability. I don't think this is a far fetched understanding of how any social network works.
Too bad they exist in a quantum superposition of publisher and platform states simultaneously, while also being neither. Pretty nice gig if you can get it!
Platform and publisher are meaningless terms born of latter day propaganda. They are about as meaningful legally as asking if they are a unicorn or a leprechaun.
