A new study of YouTube’s algorithm attracting mainstream attention this weekend claims that the online video giant “actively discourages” radicalization on the platform. And if that sounds suspect to you, it should.
The study flies in the face of everything we know about YouTube’s recommendation algorithm. There has been plenty of evidence that it pulls users down a rabbit hole of extremist content. A 2018 study of videos recommended to political viewers during the 2016 election found that an overwhelming majority were pro-Trump. Far-right allies of the authoritarian Jair Bolsonaro in Brazil say he and they wouldn’t have won election without YouTube.
Now here comes Australian coder and data scientist Mark Ledwich, who conducted this new study along with UC Berkeley researcher Anna Zaitsev. The pair looked at 768 different political channels and 23 million recommendations for their research. All of the data was pulled from a fresh account that had never viewed videos on YouTube.
In a tweet, Ledwich presents some of their findings using oddly emotive language:
“It turns out the late 2019 algorithm
*DESTROYS* conspiracy theorists, provocateurs and white identitarians.
Let’s break down why the study doesn’t measure up.
1. It’s woefully limited
The first problem: the study ignores prior versions of the algorithm. Sure, if you’re using the “late 2019” version as proof that YouTube “actively discourages” radicalization now, you may have a point. YouTube has spent the year tweaking its algorithm in response to the evidence that the platform was recommending extremist and conspiratorial content. The company publicly announced this clean-up plan early in 2019.
But in a followup tweet, Ledwich says his study “takes aim” at the New York Times, in particular tech reporter Kevin Roose, “who have been on myth-filled crusade vs social media.”
“We should start questioning the authoritative status of outlets that have soiled themselves with agendas,” Ledwich continues — ironically, after having announced an agenda of his own.
Ledwich’s problem appears to be with Roose‘s article The Making of a YouTube Radical. The story’s subject, Caleb Cain, started being radicalized by YouTube video recommendations in 2014. Therefore, nothing about the 2019 YouTube algorithm debunks this story. The barn door is open, the horse has bolted.
Cain represents countless individuals who are now subscribed to extremist or conspiracy theory-related content. Creators publishing this content have had years to get a head start. They’ve already benefited from the old recommendation algorithm in order to reach hundreds of thousands of subscribers. These channels are now popular and their content spreads due to that popularity.
Roose hit back against Ledwich in a lengthy thread:
2. It’s clearly slanted
The second problem has to do with the subjective and highly suspect way Ledwich and Zaitsev have grouped YouTube channels. He has CNN categorized as “Partisan Left,” no different than, say, left-wing YouTube news outlet The Young Turks.
The study described channels in this category as a “exclusively critical of Republicans” and “would agree with this statement: ”GOP policies are a threat to the well-being of the country.””
This is, of course, self-evidently ridiculous. CNN is a mainstream media outlet which employs many former Republican politicians and members of the Trump administration as on-air contributors. It is often criticized, most notably by one of its former anchors, for allowing these commentators to spread falsehoods unchecked.
Naming CNN as “partisan left” betrays partisanship at the root of this study.
Beyond that, there are other partisan flaws with the study such as how it groups YouTube channels from right wing partisans like Steven Crowder and Milo Yiannopoulos. The two are labeled simply as nonpartisan “provocateurs” looking to take just any position for attention. This is a blatantly false description and inaccurate grouping for the study’s two examples.
3. It doesn’t get YouTube
A third major problem: the researchers appear to not fully understand how YouTube works for regular users.
“One should note that the recommendations list provided to a user who has an account and who is logged into YouTube might differ from the list presented to this anonymous account,” the study says. “However, we do not believe that there is a drastic difference in the behavior of the algorithm.”
The researchers continue: “It would seem counter-intuitive for YouTube to apply vastly different criteria for anonymous users and users who are logged into their accounts, especially considering how complex creating such a recommendation algorithm is in the first place.”
That is an incorrect assumption. YouTube’s algorithm works by looking at what a user is watching and has watched. If you’re logged in, the YouTube algorithm has an entire history of content you’ve viewed at its disposal. Why wouldn’t it use that?
It’s not just video-watching habits that YouTube has access to, either. There are other complex factors at play. Every time you hit “subscribe” on a YouTube channel, it affects what the algorithm recommends you to watch.
Plus, since YouTube accounts are connected to a Google account, simply being logged into any of Google’s services means you’re pretty much always accumulating data for its algorithm because you’re logged into YouTube as well.
Any user can test out whether being logged in to their YouTube account matters on their own and debunk this claim. Being logged into an account versus being an anonymous user makes a major difference to the algorithm, as other researchers of YouTube radicalization have pointed out.
As experts in the field will tell you, it is extremely difficult to produce reliable, quantitative studies on YouTube recommendation radicalization for these very reasons. Every account will produce a different result based on each user’s personal viewing habits. YouTube itself would have the data necessary to effectively pursue accurate results. Ledwich does not.
We may never truly know the magnitude of YouTube radicalization. But we do know that this study completely misses the mark.