How Machine Learning Harms Us

Machine Learning

I’ve had this experience with YouTube. I discover a particular content creator who I like. And YouTube offers me another content creator who does similar stuff. So I’ll go on to watch all the videos of both these people. But then a year later, I’ll find out that there was a third content creator doing similar stuff who is similarly popular. Yet YouTube never introduced me to this content creator.

This is the essence of machine learning. If I watch a content creator’s video, YouTube is going to show me other videos from that content creator. They may show me a different content creator because they happened to have created a video exactly related to the current video I’m watching. But I’m not being offered that content creator because they are similar to the first. I’m being offered them only because they created a video that is similar to the first content creator’s video.

In other words this is one level deep.

You like this video? Here are other videos that are like that video. You like this content creator? Here are other videos from the same content creator.

Figuring out what I’m interested in in a general sense seems to be beyond what YouTube (and basically every other machine learning algorithm) is capable of. Of course, machine learning is not about providing individuals what they want. Machine learning is about looking for broad trends. In other words: what the population is doing. It isn’t about showing me videos I would like; it is about increasing overall engagement by a small amount.

The Bicycle Example

Ezra Klein has noted that his experience with machine learning indicates that it doesn’t do much. After he bought a bike, everywhere on the internet, people were trying to sell him yet another bike. And for most people it is absurd. If you just bought a bike you’re not going to buy another bike.

But I suspect that the people who have not bought a bike recently are slightly less likely to buy a bike very soon than those who have bought a bike recently. And that is what machine learning does. It makes marginal improvements to the success rate of advertising.

Machine Learning Makes Us Worse

The problem of course is that machine learning makes all of us less interesting. I have wide-ranging interests. Yet at any given time YouTube is only offering me the kinds of videos I’ve just recently watched. The whole system is designed ultimately to bore us all to death by presenting us with the same thing over and over.

And we know where this goes if you happen to be interested in politics. We’ve seen how social media with its machine learning algorithms causes people to go from conservative to fascist. Or liberal to socialist. (Not that I’m equating fascism and socialism; we could do with a good deal more socialists and no fascists.)

But the general phenomenon is that these algorithms make us far more limited and rigid. Far from opening up the world of knowledge and ideas the internet has shut them down. It’s like going into a library but only being allowed to look at anything on one shelf. But even that would be better than what we have.

The Hopeful 1980s

I first got on the internet in 1987. And it’s shocking to look back on just how optimistic I and pretty much everyone else was. It was so obviously powerful. And it seemed to offer a new way of social organizing.

There were problems of course. And those problems have been magnified exponentially. But that’s not even the worst of it. The biggest issue (by a wide margin) is the way that the commercialization of the internet has turned the very idea of free exchange of ideas on its head.

I don’t think I’m alone in craving new and interesting ideas. I think most people would go to weird and idiosyncratic websites if they knew they existed. But all the power and social forcing is on the side of a few corporations that have almost unlimited power.

There isn’t going to be a major video-sharing service that encourages people to watch new and unrelated things. That doesn’t create engagement. People watching such a thing might stop watching and, I don’t know, pick up a book. That’s not going to maximize the profits of the video-sharing platform!

Machine Learning Is Good for a Chosen Few

So in the name of making 10 cents from an hour’s worth of engagement, Google, Facebook, Twitter, and the rest are willing to help destroy liberal democracy.

And we are helpless because it’s hard to search for interesting things. It might take you 40 pages of Google results to find anything that isn’t like almost everything else. But it’s easy to just sit on Facebook and have it offer you an endless supply of just what you’ve seen before.

I’m not saying that I know what to do about it. Fundamentally I think that the human ability to find patterns has gone beyond what human nature is capable of defending itself against.

So I no longer think that humanity has much of a future. I certainly don’t think it has an optimistic future like Star Trek. But as I’ve gotten older I’ve also become more lackadaisical. There’s really nothing special about humans. And our eventual extinction really doesn’t matter that much to me.


Image cropped from Artificial Intelligence by Mike MacKenzie under CC BY 2.0.

1 thought on “How Machine Learning Harms Us

  1. Where I’ve noticed YouTube being really bad is at music suggestions. They just don’t have any! Besides the artists I’ve already watched.

    You’d think it wouldn’t be too hard to suggest music I like. If I like Charlie Parr, and I do, then fellow Minnesota bluegrass musicians Trampled By Turtles should pop up. R.E.M. should cue up The Baseball Project. Etc, etc. I wonder if YouTube is scared of getting sued by record companies for the algorithm ignoring certain acts, though, and so it won’t suggest anything new at all.

Leave a Reply

Your email address will not be published.