How ‘Straight’ is the YouTube Recommendation Algorithm?

Sorry, I disappeared into a YouTube black hole’ is something that can be heard on the weekly. In the words of the arts student Varizka (21), this means “when I’m binging one video after another just based on the recommended section, and I end up watching something weird and completely different to what I looked for in the first place.” This is a phenomenon that many users of the video platform can relate to, as the recommendation algorithm (and the ‘autoplay’ function that goes hand in hand with this) is one of YouTube’s most addictive features. The way it works is by recommending a selection of videos that is considered to be related to the video currently playing (the entry point) and, if turned on, automatically playing the first video from the list after the video from the entry point is finished playing. And yes, you guessed it: have it their way, YouTube will never let you offline.

The Black Hole That NASA Doesn’t Research – So I Did

Hopefully by now this is clear, but the black hole in this context is not the kind that is found in space. It’s produced by YouTube’s recommendation algorithm. In theory, this algorithm is supposed to recommend similar videos, but, like Varizka mentions, has a tendency to take the user on seemingly random paths through its library of content. In some cases, users have even experienced that innocent content has led them to extremist videos. This is especially true when it comes to already controversial themes, such as politics, with an example being mainstream videos about Trump surrounding the 2016 US election leading to far-right extremist content. This had me wondering: would the same be true for time old controversies, such as homosexuality within religions? By using search queries containing the key words ‘gay’ combined with each of the Big Five religions (Christianity, Islam, Judaism, Hinduism and Buddhism) respectively using a Video Network tool, I learnt that the answer to this question is a bit more nuanced than a simple yes or a no.

Visualisation of homosexuality in Christianity dataset

After having put these datasets through the data visualisation program Gephi, it was possible to identify certain trends and oddities across the different religions, like the various categories that the videos belong to. The dominant categories within four of the religions were no bigger than 30% of the entire datasets. The dataset on homosexuality in Islam, however, showed that the dominant category was news and politics, with nearly half of the videos belonging to this topic. Another interesting point taken from the categorising of these videos is how non-profits and activism scored more than 100% higher in the Christian dataset, compared to any of the other religions. Aside from these two categories, the top 8 categories also included entertainment, people and blogs, education, film and animation, how-to and style and music. With how-to and style and music being the exception (how-to and style only appeared in the top 8 categories in the Hindu dataset, and music only in the Christian and Jewish datasets), the rest of these categories appeared in similar numbers in the five religions.

To dig deeper into the datasets, I then looked at the structure of the graphs. Could they look representative of a black hole? I was surprised to see that they did. 

Visualisation of homosexuality in Islam dataset

However, after some manipulation, the networks started developing clusters and structural gaps. These are useful in identifying some of these oddities mentioned earlier. Clusters in these kinds of networks often represent extreme points in the network, as the peripheral ‘camps’ of nodes are not connected to many (if any) of the central nodes in the network. In the different datasets, there were different clusters that stood out, such as the Randy Rainbow Song Parody and Nixon-related clusters in homosexuality in Judaism, or the Queer Eye, Aligarh Muslim University and Donald Trump clusters in homosexuality in Islam. Some of the nodes may represent offensive or extreme content, such as the Donald Trump cluster which contains nodes of information on heavily right winged politics and mentions of the Swedish immigration policies which are now seemingly growing more conservative. In the same network, though, is a cluster containing nothing but videos of Tan France from the popular Netflix show Queer Eye, ranging from discussions on racism to makeovers. From homosexuality in Christianity, there seemed to more talks, educational talks such as TedX-talks and Google talks investigating the relationship between religion and sexuality, as well as religious talks from the anti-LGBT+ pastor John McArthur. This was a noticeable trend across all religions included in the project. Within the dataset on homosexuality in Hinduism, both videos on LGBT+ persecution in India appeared, as well as videos of Indian gay proposals. For homosexuality in Buddhism, clusters of Buddhist gay marriage related content, as well as educational videos on how to give up on sexuality all together in order to live according to Buddhist ways.

“I once went on YouTube to for a makeup tutorial, and then 2 hours later I was sitting watching conspiracy theories about the industrial revolution”

Varizka Anjani

In other words, it seemed as though, yes, there were extremist directions that the algorithm could lead you, but these extremes represent a selection of information, ranging from liberal to conservative, from factual to fake, from gay to straight. The main body of the networks, containing the most and most connected nodes, also visualised this, with their many camps of ideas merged into larger networks. In the case of homosexuality in different religions, it seems that the extreme points that the YouTube algorithm might take its users to merely reflect the variety of opinions there are within the topic on a larger scale.

To conclude, here are some navigation tips for navigating the black hole:

  1. Turn off Autoplay
  2. Be aware of your camp
  3. Look out for the peripheries
  4. Curiosity can mislead you