A decade ago, the writer Eli Pariser popularized the term “filter bubbles,” which refers to the idea that search and social media algorithms wrap individuals in information bubbles that are tailored to their interests and behaviors rather than ones filtered by traditional gatekeepers like journalists. However, academic research has largely failed to support Pariser’s thesis, suggesting that filter bubbles may not be as pervasive as feared.
Yet in the debate over algorithms’ impacts on society, the focus on filter bubbles may be something of a red herring. While personalized filter bubbles might not be the most important problem, we show in our research that even general recommendation algorithms remain highly problematic. Even when they are not personalized, recommendation algorithms can still learn to promote radical and extremist content.
YouTube’s channel ecosystem
Our research focused on how YouTube’s algorithms “perceive” the platform’s ecosystem of channels—which is to say, how the algorithms connect different channels to one another through recommendations about which channels to follow. To measure how YouTube’s channel recommendation algorithm structures those channels, we collected lists for political channels and the most popular mainstream channels in Germany and in the United States. Next, for each channel we followed YouTube’s channel recommendations for three steps, which resulted in a channel network with 8,000 channels in the German case and 13,529 channels in the United States. As YouTube does not personalize channel recommendations, those recommendations were relatively stable and can give a good idea of how its algorithms structure the channels on the platform. By contrast, video recommendations are personalized based on a user’s viewing habits and are notoriously unstable and even harder to generalize.
In our recently published paper, we sought to understand what factors contribute to the decisions made by YouTube’s channel recommendation system. Based on the academic literature, we assumed that YouTube’s algorithm would not make channel recommendations at random, but would instead base its recommendations on the themes and language of each channel, as well as its creator’s location. Our results confirmed our assumptions: In both the German and American cases, YouTube’s recommendation network differed significantly from what we would expect if it made recommendations at random.
More importantly, our results also confirmed that YouTube’s channel recommendations form around themes. Whether for sports, music or politics, YouTube had created thematically consistent communities that formed around prominent channels, such as PewDiePie for gaming or Worldstarhiphop for hip hop. This held in both the United States and Germany.
In the political realm, we manually identified three general communities on U.S. YouTube: a relatively small “Far-left” community, a bigger “Liberal/mainstream news” community, and a big “Far-right & politics” community. Within the “Liberal/mainstream news” community, The Young Turks and CNN were the most recommended channels. Within the “Far-right & politics” community the most recommended channels were The Alex Jones Channel (the notorious, now banned far-right conspiracy channel), Styxhexenhammer666 (a far-right libertarian who has questioned the Holocaust), and only then Fox News.
In the German context, we were only able to identify one political community. Though this community included establishment political parties like CDU or SPD, the channels most recommended by YouTube’s algorithm were Der Volkslehrer (a now banned far-right extremist channel that featured Holocaust denial), RT Deutsch (the Russian state-controlled German-language propaganda channel that often features far-right speakers), and a channel from the far-right extremist political party AfD.
Given that the filter bubble hypothesis posits that most channel recommendations would stay within a given community, we also sought to understand the heterogeneity of YouTube’s communities. To do so, we calculated the E/I index—which measures the external versus internal links of a given community—of the communities we identified. While random networks were more outward-oriented, YouTube’s networks for both Germany and the United States were highly homophilous, meaning that most recommendations from YouTube’s recommendation algorithm stayed within each community. And while this doesn’t mean that YouTube’s algorithms lead to filter bubbles, we posit that these highly homophilous communities are a prerequisite for the formation of filter bubbles.
Community by algorithm
YouTube’s algorithms work really well at detecting shared themes and forming communities around them. Yet these algorithmic decisions lack nuance: They cannot distinguish between “news” on the one hand and “political punditry” on the other. And without the ability to make nuanced determinations about content, the algorithms sidestep questions about the veracity of the information presented and extremist speech.
YouTube’s algorithms also appear to reward especially successful channels within a given community with more recommendations, making already successful channels even more successful. We refer to this as a “digital Matthew effect.” In practice, this meant that on any given political channel, there was a significant chance that one of the recommended channels was Alex Jones. It wasn’t just a handful of YouTube channels that recommended Alex Jones: More than 160 other channels in our network recommended the conspiracy theorist. And in doing so YouTube not only lead more people to Alex Jones but also contributed to the normalization of radical content.
The ability to disrupt the Matthew effect is one way YouTube holds creators to account. During the course of our research, the popular YouTuber Logan Paul came under fire after he uploaded a series of videos from his travels to Japan. These ranged from racist to offensive and caused a huge outcry. In prior iterations of our U.S. network, Paul was one of the most recommended channels, but after the Japan incident he seemed to disappear from our network. But that was not really what happened: Paul was still in our network, he just didn’t receive as many recommendations as before. And while we do not know exactly what happened, a likely explanation is that YouTube manually downranked the YouTuber as a punishment.
The homophilous communities created by YouTube around shared themes, locations, and languages lack nuance and, especially in the political context, are more extreme than the general mass media landscape. This is only partly YouTube’s doing. YouTube is more extreme in part because more extreme people use YouTube to create political content that is not featured in the mainstream media. And while the mainstream media rarely creates content specifically for YouTube and its community, political YouTubers do. They monitor and jump on trends, speak out on YouTube-specific issues, engage with the users who write comments, do live streams on YouTube, appear on each other’s shows, etc.
But people that come to YouTube to watch the news and get information rarely understand the differences between mainstream outlets and YouTube. Nor do they know that YouTube’s political sphere leans heavily to the right, nor that for YouTube’s algorithm there is no distinction between a conservative and a Holocaust denier. We argue that this can lead to a “digital Thomas theorem”: While users may or may not be influenced by YouTube’s recommendations, they nevertheless will take note of the topics that are being talked about and the channels that they see. This, in turn, might lead to a normalization of more radical content and, in some cases to the radicalization of individuals.
Finally, the algorithms may also be leading to real-world harms. While we have no way of perfectly measuring the impact of YouTube’s channel recommendations, we have conducted a separate analysis of YouTube comments in the far-right community in Germany. There, we were able to detect that the community grew more centralized with stronger overlaps between the audiences of different far-right channels and, thus, more connected over time. This happened especially around topics such as refugees. YouTube’s algorithm is one possible culprit that may have accelerated this process and contributed to the group of channels’ shared sense of community. But we cannot estimate the real impact of the algorithm over time.
Indeed, it is unclear what YouTube’s algorithms cause and do not cause. The academic research on its algorithms remains limited and hampered both by the massive amount of data that content creators and users on YouTube produce and by YouTube itself. Access to data is limited and data about traffic generated by recommendations is only available to channel owners. When conducting research together with journalists from The New York Times on YouTube in Brazil, we stumbled upon a community of videos that featured barely clothed children. When alerted to this fact by the Times journalists, YouTube acted swiftly and implemented changes to turn off the recommendation system for videos that featured children. At the same time, YouTube also deactivated its channel recommendation system because it was allegedly barely used. We had used it to identify the channels but also for this work that we are presenting here. So while we still cannot say that YouTube’s recommendation algorithms cause a specific response, we found evidence that the algorithms nevertheless can cause harm.
A way forward
In this context, we suggest that YouTube should deactivate recommendations for political content. It is clear that YouTube can’t distinguish between different types of content and will as a result recommend radical content. It is also clear that YouTube’s algorithms will always be prone to recommend radical content due to the userbase on the platform and what Michael Golebiewski and dana boyd call “data voids”, that in the absence of quality data algorithms that are charged with generating recommendations will recommend what is available, even if that means recommending radical content and disinformation. These data voids are exploited by the more extreme fringe creators on YouTube. We have already witnessed YouTube shutting down its recommendation algorithms when facing pressure and realizing that they were causing real harm; it is time it do the same for its recommendation systems for political content.
Jonas Kaiser is an assistant professor at Suffolk University, faculty associate at the Berkman Klein Center for Internet & Society at Harvard University, and associate researcher at Humboldt Institute for Internet & Society.
Adrian Rauchfleisch is an assistant professor at the Graduate Institute of Journalism at the National Taiwan University.
Google provides financial support to the Brookings Institution, a nonprofit organization devoted to rigorous, independent, in-depth public policy research.