Manoel Horta Ribeiro (@manoelribeiro),
In this blog post I will make some comments about a recent pre-print on YouTube Radicalization, which brought some new and exciting ideas about radicalization on YouTube. The paper is named “A Supply and Demand Framework for YouTube Politics”. I believe that the authors have nice arguments in their criticism, but also that the framework they propose has their own shortcomings. To be fair with them, I think they make it quite clear in the paper that they do not wish their framework to be a “final answer”. In that sense, I believe that this blog-post is exactly what their paper aimed to achieve: “to encourage a broader scholarly analysis by pointing out that the algorithm is just one affordance of YouTube”. Okay maybe calling my blog “scholarly analysis” is too much, but I think you got my point.
I would first like to thank the authors for the (working) paper. I think it is really interesting, I learned a lot of stuff about media studies on YouTube that I did not know about, and it just felt like a fresh perspective overall. The paper brings interesting criticism to what the authors call *The Zombie Bite” Theory of YouTube, which is that:
YouTube audiences are at risk of far-right radicalization and this is because the YouTube algorithm that was designed to maximize the company’s profits via increased audience time on the platform has learned to show people far-right videos.
The authors then propose an interesting framework, which they name “Supply and Demand”, where, according to the authors:
(…) the novel and disturbing fact of people consuming white nationalist video media was not caused by the supply of this media “radicalizing” an otherwise moderate audience. Rather, the audience already existed, but they were constrained by the scope of the ideology of extant media. The expanded supply allowed them to switch into consuming media more consistent with their ideal points.
I think that Munger and Phillips touch a key issue with the narrative they name “The Zombie Bite Theory”, namely that this idea that “a model of YouTube media effects that centers the recommendation engine is implausible, an unfortunate update of the hypodermic needle model of media effects that enjoyed some prominence in the 1930s and 1940s but which has been consistently discredited ever since”. Indeed, I agree this is too simplistic, and that the way this is attractive to the media probably makes it easier to pursue this narrative. Even in our paper, where the findings where much more related to the migration of users than to the algorithm, we consistently got framed as a study about “The Algorithm”.
I believe that the big issue here is that there are three mechanisms at play. The first one is radicalization itself, a phenomena that has happened in the past, long before recommender systems existed. The second, and this is what is captured best by the “Suply and Demand” framework, is that there are significant differences in how YouTube content is created and consumed, which favours niche ideologies (I honestly found their arguments on this fascinating, perhaps because they understand much more about media than I do). The third is this new factor, the algorithm, which so far has been pretty hard to study quantitatively.
Interestingly enough, it is hard to tell what is the influence of each of these mechanisms in the surge of fringe content on YouTube: is it merely a matter of the political times we live in? Is it because of how this new media operates? Or is the algorithm “infecting” users with its zombie-esque bites? I don’t know.
I have now placed their framework as mostly answering one out of three questions that we are faced (the one which I called the second mechanism). Yet, my belief that merely their framework is not enough to explain the whole scenario goes slightly beyond this mere intuition. I believe that their “Supply and Demand” approach is unable to explain the phenomena that we have the most evidence about: user migration. Somehow, YouTube has become a fertile ground, not only for the Alt-right, but for pattern of content consumption where users increasingly consume more extreme content. In that sense, is does not seem to me that a latent bunch of white supremacists suddenly found the content they strived for, but that they found in the platform a place to “draw in” new individuals into their fringe ideas.
In the paragraphs below, I make some points w.r.t. their quantitative analysis and to their comments on our paper.
Comments about their quantitative analysis: I see two main shortcommings in their quant. analysis: 1) It is unclear to me if the downfall of videos is due to the significant banning of Alt-right channels on YouTube and the migration of such content to other streaming services such as BitChute. For example, in the Alt-right, the channels ranked 1st, 3rd, 5th in terms of views were banned. Would be interesting if they compared their sample with mine :). 2) Also, they evaluate the search in previous years by tinkering the search result of the YouTube API. I think there is no clear reason to believe that the historical results of today’s API reflect in any way the search engine in the past. If there are filters for topics that have attracted polemic videos (such as the ones they searched), it seems quite likely that these would be in the current API when you search for past videos.
Comments about comments about our paper: Indeed, as the authors have noted, our paper fails to demonstrate that the algorithm has had a noteworthy effect on the audience for Alt-right content. Yet, this should be taken as a failure to provide positive evidence towards the Zombie Bite theory, and not as a evidence that the algorithm does not work. This is because (and that is stressed in the paper): 1) The algorithm was probably different in previous years, and we only looked in a couple of months of 2019. In fact, Google is allegedly trying to prevent this content to surface. 2) We don’t have personalization. Another point that is worth making here is that they complain about our usage of “Related Channels”. Yeah, indeed it does not play such a big role in YouTube, but, here we may find some other cues on how the recommendation system works for the “cold start” case. While many of the recommendations from the video recommender system simply point at popular channels (for example Fox News or CNN), here we actually get recommendations that are much more aligned topically. Indeed, we did not provide this context, and I will probably make this more clear in the next version of the paper.
Note: In this month still I am drafting another blogpost about some of the criticism and the attention our paper received. It has been hard to write much these days as I just recently moved :)