A while back when I was streaming, we got onto the topic of math, and one of my stream viewers mentioned the Numberphile videos on Youtube. Now, I had never watched a Numberphile video because I didn't expect them to provide anything new to me. But why not give them a chance? So I watched one.
It was pretty much as I expected. The video had a reasonable topic, but the presentation lacked any real substance and didn't answer any of the interesting questions. But maybe I just hit a bad one. So I watched another. Wasn't any better. Tried a third time. It wasn't long before I gave up and decided that Numberphile just straight up isn't going to be good.
Soon after that I noticed that Youtube was recommending me Numberphile videos. Of course they would, because from their perspective I went from one video on the channel to watching multiple. That behavior looks identical from their perspective regardless of whether I'm actually enjoying the videos or trying to find any redeeming quality to videos that are generally popular but seem terrible to me.
Youtube does allow you to indicate that you're "not interested" in one of their recommendations, and it does seem to help if you keep doing it, at least a little bit. But that's still extra effort required to get rid of videos you don't want to see. How long would it take for Youtube to stop giving those recommendations without me ever making an explicit rejection? I don't know but I bet it would be a long time, especially since I don't use Youtube all that often.
When I, as a normal user, go to Youtube, I am looking for videos that I will enjoy. On the other hand, the content creators are not rewarded based on my enjoyment. Rather, they are rewarded based on whether or not I open the video in the first place. Thirdly, while the recommendation engine is presumably trying to maximize my enjoyment, measuring enjoyment is very hard, so I would bet that it uses me opening the video as a proxy for that enjoyment. So there's this disconnect between what's being measured and the real goal, which means that sometimes it will be spectacularly wrong, like in this case.
The "not interested" feedback does help out with cases like this where I can accurately predict that I will not want to watch the video. What about the less clear cut cases. Maybe a channel is 50-50 on videos I enjoy versus videos that I don't enjoy. When the recommendation engine suggests a new video from this channel, then I can't reject it without opening it, but after I open the video, how do I inform the recommendation system that it failed?
This situation closely parallels the situation with news stories. Just like Youtube videos, news stories are not rewarded by how much information a person actually got out of it, but by how many people vieweed that story. There's very little short-term incentive in this case to make an article that is actually interesting for the majority of readers. On the contrary, the majority of their rewards are simply dependent on getting people to open the story. The result? Clickbait.
These are examples of a more general pattern, where what is measured is what gets optimized, regardless of what the intention of the measurement was. Optimizing the pass rate of students doesn't necessarily cause students to become better educated, it might instead mean that the tests became easier. Optimizing the number of accounts on a website might lead to lots of people who create an account and the never come back. So to fix the problem of clickbait, the best solution would be to find a way to actually measure how much value the users get out of watching a video, and then reward based on that. Proxying value with view counts, or even total time viewed, is too easily gamed.