It was probably a month before The New York Times posted its article On YouTube Kids, Startling Videos Slip Past Filters that I noticed three things while visiting my parent’s house when they were babysitting my nephew: first, that he always had the iPad — or ‘Pad-Pad’, as he calls it; second, that he was always on YouTube; third, that he was really good at finding every possible ‘Family Fingers’ video on the site. Turns out the videos were better at finding him.
Since that Times piece there’s been a lot said about the problems with YouTube and its ‘algorithm’, as it’s broadly referred to, and whether or not that algorithm — or the exploitation thereof — is or has endangered children in particular. NPR has a wonderful piece on it, and in it links to James Bridle’s harrowing blog post appropriately titled Something Is Wrong on the Internet.
And something is wrong, at least on YouTube. And it’s probably not insidious in nature; rather, it’s precisely human nature and the keen observation of people who see profits that I believe have lead us to this point in time, where YouTube must accept that SEO and neural networks aren’t going to sweep away the kind of grotesque material that was reaching children, nor are the viewing habits of children going to change any time soon.
The problem with YouTube is that it’s popular. Like, really popular. Like, 300 hours of content uploaded a minute popular. Being, as they are, swamped with content, YouTube must deliver the best videos to keep viewers viewing without making them bog through trash. Enter an algorithm that prioritizes getting users to sit and watch content they care about it in its entirety, preferably back to back. I think it’s fair to say they’ve succeeded, as YouTube, in many ways, has become the new television. I have YouTube on in the background as I write this.
The problem with parents is that their lives are very busy, and their children need to be preoccupied. Enter the combination of smart devices with YouTube on them.
The problem with children, and especially toddlers, is that they will sit and watch. And watch. And watch.
After researching how the algorithm behind YouTube’s many systems work, I have decided that their SEO algorithm and its explicitly stated goals are rather incredible and founded on reasonable knowledge and assumptions about the viewing habits of users. I don’t believe their algorithm has been exploited. Instead, I believe their users have been, and more specifically, children.
Before we give a cursory explanation of YouTube’s Search and Discovery system — the algorithm that decides which content a user is most likely to engage with and be satisfied by — I would like to wager that had harmful material reaching children not been the lead of these stories, children would still be exploited on YouTube today. The people responsible for making harmful content were surely a small percentage of the people exploiting the viewing habits of children in order to make money. One angle that I am not covering so much is the role of advertising in all of this, but I have the sad feeling that people are less upset with the fact that children are really great and easy to advertise to, and that that’s what really lead to harmful material eventually being disseminated to them.
How YouTube’s Algorithm Used to Work
YouTube’s early incarnation of its ‘Search and Discovery’ algorithm equated clicks with views: if a user clicked play, and the video player was loaded, then these clicks counted as views. It wasn’t concerned with how long a video was watched, simply that a user chose to start watching it. This lead to multiple different problems, including misleading thumbnails or descriptions for videos all with the intent to get users to click. More or less, if a creator could get you to click, they got your ‘view’.
It wasn’t until 2012 that the algorithm was revised with the intent to promote and track user engagement and satisfaction, rather than just clicks. This was done via ‘Watch Time’ and ‘Watch Sessions’.
What Is Watch Time and How Does It Work?
Matt Gielen of Tubefilter wrote a very insightful piece on Watch Time that I highly recommend. In it, he covers how YouTube has emphasized the importance of Watch Time over views. YouTube’s stated goal on its Creator Academy page is to ‘Help viewers find the videos they want to watch, and maximize long-term viewer engagement and satisfaction’. This last part is key: in order to provide users with videos that will hopefully continue to engage them, whether they are from the same creator or not, YouTube had to determine how user engagement can be quantified and used to inform their algorithm.
The way YouTube arrives at the value for a given Watch Time is simple: take a creator’s total number of views and multiply them by the Average View Duration for that creator. As Matt Gielen points out, however, that metric isn’t incredibly useful for the creators themselves, as they “can’t impact, change, or optimize for [their] ‘Watch Time’ without impacting, changing, or optimizing for its components.” I believe there’s a couple reasons why YouTube prioritizes Watch Time above other metrics: one, it’s very hard to manipulate, or exploit; two, along those same lines, it’s both transparent and opaque. In other words, even though a creator has access to this information, the information itself doesn’t necessarily expose vulnerabilities to how the algorithm works.
Matt Gielen goes into even greater depth about Watch Time, but with the caveat that much of the information isn’t coming from fully credible sources; or, at least, the information isn’t coming from YouTube itself. The main takeaway, however, is that Watch Time delivers YouTube a more useful metric for what users are actually watching, and for how long. Again, Watch Time is a metric for engagement.
What About Watch Sessions?
The second part of YouTube’s stated goal to engage and satisfy viewers is to ensure that the videos users are watching leads them to watch other videos, regardless of if those videos are coming from the same creators. I would refer to this metric as retention. After doing more research on this particular facet of YouTube’s algorithm, I have yet to be able to find conclusive information on how this metric is measured and used. Suffice it to say, however, that YouTube has a greater vested interest in its entire platform more so than for individual users.
A Quick Note on Other Metrics
So far I’ve merely given a broad overview of YouTube’s algorithm. In 2016 Google released a scholarly article detailing YouTube’s Recommendations and the system of neural networks that determine what content to recommend to individual users. It’s concise, insightful, and complicated.
We could cut a slice out of YouTube and find ways in which its algorithm functions differently, whether it’s currently delivering recommended videos, trending videos, recommended videos from a given channel, multiple channels but a given topic (gaming, sports, cooking, etc.), and on and on. It’s my belief that across the platform the end goals are the same: engage and retain viewers.
One metric that I would have thought was more important was user ratings for videos. In my research, I only saw it mentioned a few times, and usually weighted near last. This surprised me, given how often culture blogs and websites focus on instances of overwhelmingly positive or negative responses to a given video based on user ratings. The first example that came to mind was the 2016 Ghostbuster’s trailer.
Quickly it became clear how useless this metric probably is: of the 1 million dislikes the trailer garnered, it had an overall total of 44 million views. That means that only .02 percent of viewers didn’t like the content. This probably tells us more about the rating habits of users than it does their genuine viewing desires. (And maybe their gender politics, but that’s a different depressing topic.)
I want to further research and understand the systems behind YouTube writ large, but for now, it’s worth noting that any company dealing with the sheer volume of content that YouTube attempts to curate on a daily basis is going to necessitate an equally large amount of data to optimize their algorithms to meet individual user’s needs. To quote Ric Mazereeuw’s quick summary of Google’s scholarly article on the matter, “the key thing to remember is that YouTube isn’t in the business of judging whether your video is ‘good’ or not.” Rather, their algorithm prioritizes user interaction with videos and artificial intelligence that can learn and respond to the data collected to continue delivering users content they think is good.
User Interaction and YouTube in the Age of iPad Toddlers
My main takeaway from my research is that YouTube’s curation systems work as designed. The problem is that so do children and toddlers. The ideal end game for the feedback collected from user interaction with content would be for the user to never realize that the algorithm is actually working. This works best when the user is at least cognizant of their interactions with YouTube to some degree. Children and toddlers are less likely to be cognizant of what those interactions really mean.
The true issue with retention and engagement is that unattended younger users are more likely to remain engaged and continue watching regardless of any standard of quality or appropriateness. And again, YouTube isn’t interested in the quality of the content; rather, they’re collecting data precisely to determine what you, the user, thinks is quality. What this necessitates is a self curating system. What this necessitates is algorithms. However, the algorithm behind YouTube’s content delivery and search systems doesn’t seem to account for children. With no real quality control or curation, the system trusts the user to deliver it worthwhile data in order to deliver them worthwhile content. Somewhere in this symbiotic relationship something failed.
A Final Note
When it comes to content curation that has an added layer of protection for our most vulnerable users, I don’t think algorithms are necessarily the answer; however, when faced with mass quantities of data and content to curate, algorithms are the answer that’s necessitated. However, as our algorithms become more complex and our quantities of data and volume of users increases exponentially, a company like Google must also accept that their platforms are reaching younger and younger audiences who aren’t giving them the kind of data they might be expecting. At best, there’s a target demographic being underrepresented and under-attended to, but oversaturated in the amount of trash content bombarding them; at worst, the system is working as planned.