Hacker News Reader

> Why wouldn't we expect these systems to learn that among other things, information that increases polarization results in the desired goal of increasing revenue, absent of any deliberate intervention by human beings?

This is all well and good, but revenue is not an input to most machine learning algorithms at Google. First off, there is a pretty serious firewall between search and ads, so any organic results will be completely oblivious to revenue. Search systems literally do not have access to that kind of information. Even within ads, the concept makes sense only if you don't think about it too much. The big money-maker ads are when someone is trying to buy something. Ads around viral content are just not something that people can pay much for, because the margins in the news business are very thin.

Now if you want to make an article that the media has learned that spectacle and outrage drive clicks, I am here for it. (But this has always been true. Yellow journalism was a thing more than a century ago.) And I guess you can make an argument that Google facilitates this by pointing you to the media, and by participating in the feedback loop. But I don't think Google revenue is as causally linked to outrage as you seem to believe. After all, correlation usually doesn't indicate causation.

I might be wrong but I was definitely told by a YouTube rep that videos that makes more revenue get promoted more on the platform. It's not the only signal, but it is a signal.

My thought is that there's a correlation between attention and a viewer's willingness to spend money that can be seen without directly looking at revenue. For instance, letting a music playlist run would look a lot different than watching a video I'm actively paying attention to. Or a detailed discussion on credit card churning would probably be more niche than a clip from SNL. But former would probably have a more engaged audience and be more willing to act on an advertisement for a credit card. So if you're optimizing for user engagement, you'll probably increase profit.

It rose after cable deregulation, which gave us 500 channels and dozens of 24/7 outrage outlets.

There was plenty of outrage going around on Fox and Rush Limbaugh talk radio before Google. Remember, Roger Ailes pretty much planned for a lot of this.

Since neural networks train using the data of the masses I would exactly expect the effect to funnel towards consensus as opposed towards controversy. Let autoplay run long enough and eventually we all get the same videos.

I think there's plenty other explanations for the current political climate. It has perhaps moreso to do with how the media was previously controlled by small number of big players versus the many smaller players we have today with more extreme views.

Yeah, the ironic part is that roi for advertisers on google search is low to negative. Not only does google sell its keywords, thus perverting the index itself, they also fleece the advertisers who seem to do buys solely to "keep up with the joneses". Most actually results-based advertising will avoid organic search buys; but most digital marketing companies are not results-based.

I use ad blockers so I don't how this theory would explain people with ad blockers that are political.

Neural network is bias because of the data. It's always the data. gfycat has an article about this. They have a huge user base that's into kpop and their training dataset didn't have much Asian people in it so it kept on identifying kpop star as women.

Ad blockers don't help against Youtube recommendations and autoplaying next recommended video.

Studies have documented that if you let YouTube auto play for a while, the videos would becomes increasingly more radical.

I hear this a lot - but are people really watching longform political YouTube videos and passively letting the algorithm choose their next video for extended periods of time? I very rarely watch a political video on YouTube but when I do it's usually linked from a site like Reddit, and then I'm very unlikely to watch the next video in the queue after it finishes.

If the algorithm actually does suggest more and more radical videos I wonder if it's a case of reverse causation - the algorithm suggests a more radical video because it learns what people actually want to do and choose to do, which is watch the most outrageous and/or bloodcurdling video out of the options presented to them?

> Ad blockers don't help against Youtube recommendations and autoplaying next recommended video.

They actually do!

Doesn't help with autoplay, but adblock lists to get rid of recommendations on YouTube: https://youtube.adblockplus.me/

I think they changed this or are working on changing it.

YouTube recommendations used to go towards more popular, meaning all playtime eventually went to Gangnam Style. So they went the other direction, and gave more view time to radical/fringe YouTubers, which they're getting flak for.

That sounds super interesting, can you share a link?

Several articles.

https://www.google.com/search?q=youtube+recommendations+more...

I would say it is the rise of digital media. Newspapers had a captive audience, now in the digital world everyone has to compete for attention.

Google rise happened at the same time as digital media - both coinciding with rise of internet.

There's an entire research field that works and publishes on this.

Google/FB etc makes a lot of money for (mostly) liberal people by making as many people as possible non-liberal. In the future the ML's sole job will keep that momentum going but will be trained on perfecting _exact_ ratios of polarization that's needed to prevent a civil war. In other words fine tuning where that tipping point maximized profit at an acceptable level of violence.