[FullstakBlogger]

Two questions:

1. Whose role is it to make judgements about whether people should be able to find certain kinds of content?

2. Do you think modern recommendation systems are better at keeping people away from that kind of content?

The thing is, as of now, if a recommendation algorithm identifies you as someone that responds to that kind of content, whether negatively or positively, it'll show you so much of it that you think it's the entire world. Even if you want to see something else, you have to have the discipline to "train the algorithm" over days, weeks or months to break free.

I feel that any suggestion things are better now is just impossible to believe.

[IIAOPSW]

>1. Whose role is it to make judgements about whether people should be able to find certain kinds of content?

Well for starters, there's the user. Once youtube decides you like a certain type of content that it wants to amplify, there is no "ok enough of this" button. You can't dismiss the categories it makes up for you.

And for enders, well your real question comes down to the question of if sufficiently effective psychological engineering contradicts, at least for some large fraction of people, personal agency. Suppose there was an adversarial patch we discovered that you could hide in a banner ad to make 15% of people read a completely different article than the one that is actually in front of them. That's may not be a highly efficient hack, but its more than the margins on nearly any election. How do we police these newly discovered neural-network hacks?

There's no question in my hypothetical on the matter of the victims personal agency in forming opinions and seeking information. I took care of that in the premise. They have no control, its a design flaw in the brain that is being exploited. But, where along the continuum of plausible technology does one draw the line between brains being hacked and people having agency in what they choose to see and believe?

I don't know either.

[renegat0x0]

Huh. Recommendation system is based around how much addicted you are. Therefore clickbait videos, or reaction videos will have better recommendation than a normal video.

The second goal is to make you slightly more angry, so that it is more probable that you will engage with the video.

The third is that you will watch more advertisements the more you watch. It is not in the best interest of the algorithm to give you answer right away.

It also does not have to be a bad will of YouTube/Google. After all you train neural networks to have desired outcome. The goal is to capture users attention.

[eddythompson80]

> 1. Whose role is it to make judgements about whether people should be able to find certain kinds of content?

General public pressure that ultimately comes down to money (a.k.a: advertisers) which is controlled by general social pressure. With the previous rabbit hole effect, it's trivial for a reporter or an activist or anyone to record a 5 minute video where you start with a debate on a news channel or some very tame video about [Insert undesirable topic] then through 1 click on the top/next recommended video find yourself in an endless rabbit hole of full [Insert undesirable topic] videos. Then publicly call/shame those advertisers. Advertisers don't want associate their brand with that. If YouTube wants to be a bastion of free speech absolutism, that's their prerogative. Maybe it's good for their brand. We just want to sell Coca-cola or shoes or cars or whatever. To score brownie points with the public, we will pull our ads until YouTube gives us assurances that they are not promoting that [Insert undesirable topic].

Without the big advertisers, YouTube goes bankrupt in few months. Just like any traditional media company YouTube answers to their advertisers.

> 2. Do you think modern recommendation systems are better at keeping people away from that kind of content? > The thing is, as of now, if a recommendation algorithm identifies you as someone that responds to that kind of content, whether negatively or positively, it'll show you so much of it that you think it's the entire world.

That's a false dichotomy. Recommendation systems are whatever we make them to be. For example, maybe a recommendation system that recommends you content based solely on *your* own preference + the current content you're consuming, is more likely to exhibit rabbit hole effect characteristics. One that takes the subject matter into account and evaluates it equally against some other criteria might have a different characteristics.

Ultimately it's a difficult problem to answer because it's a social problem, not a technical one. Just because you're a video or a social media or generative AI platform built on tech, doesn't absolve you from social responsibility if you're large enough to have wide spread social impact. That was the case with traditional media company. The fact that you're designing and creating algorithms and statistical models to recommend, promote, or generate content means you're on the hook for what those models produce.