The World Will Never Know the Truth About YouTube’s Rabbit Holes

The site’s crackdown on radicalization seems to have worked. But the world will never know what was happening before that.

Around the time of the 2016 election, YouTube became known as a home to the rising alt-right and to massively popular conspiracy theorists. The Google-owned site had more than 1 billion users and was playing host to charismatic personalities who had developed intimate relationships with their audiences, potentially making it a powerful vector for political influence. At the time, Alex Jones’s channel, Infowars, had more than 2 million subscribers. And YouTube’s recommendation algorithm, which accounted for the majority of what people watched on the platform, looked to be pulling people deeper and deeper into dangerous delusions.

The process of “falling down the rabbit hole” was memorably illustrated by personal accounts of people who had ended up on strange paths into the dark heart of the platform, where they were intrigued and then convinced by extremist rhetoric—an interest in critiques of feminism could lead to men’s rights and then white supremacy and then calls for violence. Most troubling is that a person who was not necessarily looking for extreme content could end up watching it because the algorithm noticed a whisper of something in their previous choices. It could exacerbate a person’s worst impulses and take them to a place they wouldn’t have chosen, but would have trouble getting out of.

Just how big a rabbit-hole problem YouTube had wasn’t quite clear, and the company denied it had one at all even as it was making changes to address the criticisms. In early 2019, YouTube announced tweaks to its recommendation system with the goal of dramatically reducing the promotion of “harmful misinformation” and “borderline content” (the kinds of videos that were almost extreme enough to remove, but not quite). At the same time, it also went on a demonetizing spree, blocking shared-ad-revenue programs for YouTube creators who disobeyed its policies about hate speech.Whatever else YouTube continued to allow on its site, the idea was that the rabbit hole would be filled in.

A new peer-reviewed study, published today in Science Advances, suggests that YouTube’s 2019 update worked. The research team was led by Brendan Nyhan, a government professor at Dartmouth who studies polarization in the context of the internet. Nyhan and his co-authors surveyed 1,181 people about their existing political attitudes and then used a custom browser extension to monitor all of their YouTube activity and recommendations for a period of several months at the end of 2020. It found that extremist videos were watched by only 6 percent of participants. Of those people, the majority had deliberately subscribed to at least one extremist channel, meaning that they hadn’t been pushed there by the algorithm. Further, these people were often coming to extremist videos from external links instead of from within YouTube.

These viewing patterns showed no evidence of a rabbit-hole process as it’s typically imagined: Rather than naive users suddenly and unwittingly finding themselves funneled toward hateful content, “we see people with very high levels of gender and racial resentment seeking this content out,” Nyhan told me. That people are primarily viewing extremist content through subscriptions and external links is something “only [this team has] been able to capture, because of the method,” says Manoel Horta Ribeiro, a researcher at the Swiss Federal Institute of Technology Lausanne who wasn’t involved in the study. Whereas many previous studies of the YouTube rabbit hole have had to use bots to simulate the experience of navigating YouTube’s recommendations—by clicking mindlessly on the next suggested video over and over and over—this is the first that obtained such granular data on real, human behavior.

The study does have an unavoidable flaw: It cannot account for anything that happened on YouTube before the data were collected, in 2020. “It may be the case that the susceptible population was already radicalized during YouTube’s pre-2019 era,” as Nyhan and his co-authors explain in the paper. Extremist content does still exist on YouTube, after all, and some people do still watch it. So there’s a chicken-and-egg dilemma: Which came first, the extremist who watches videos on YouTube, or the YouTuber who encounters extremist content there?

Examining today’s YouTube to try to understand the YouTube of several years ago is, to deploy another metaphor, “a little bit ‘apples and oranges,’” Jonas Kaiser, a researcher at Harvard’s Berkman Klein Center for Internet and Society who wasn’t involved in the study, told me. Though he considers it a solid study, he said he also recognizes the difficulty of learning much about a platform’s past by looking at one sample of users from its present. This was also a significant issue with a collection of new studies about Facebook’s role in political polarization, which were published last month (Nyhan worked on one of them). Those studies demonstrated that, although echo chambers on Facebook do exist, they don’t have major effects on people’s political attitudes today. But they couldn’t demonstrate whether the echo chambers had already had those effects long before the study.

The new research is still important, in part because it proposes a specific, technical definition of rabbit hole. The term has been used in different ways in common speech and even in academic research. Nyhan’s team defined a “rabbit hole event” as one in which a person follows a recommendation to get to a more extreme type of video than they were previously watching. They can’t have been subscribing to the channel they end up on, or to similarly extreme channels, before the recommendation pushed them. This mechanism wasn’t common in their findings at all. They saw it act on only 1 percent of participants, accounting for only 0.002 percent of all views of extremist-channel videos.

This is great to know. But, again, it doesn’t mean that rabbit holes, as the team defined them, weren’t at one point a bigger problem. It’s just a good indication that they seem to be rare right now. Why did it take so long to go looking for the rabbit holes? “It’s a shame we didn’t catch them on both sides of the change,” Nyhan acknowledged. “That would have been ideal.” But it took time to build the browser extension (which is now open source, so it can be used by other researchers), and it also took time to come up with a whole bunch of money. Nyhan estimated that the study received about $100,000 in funding, but an additional National Science Foundation grant that went to a separate team that built the browser extension was huge—almost $500,000.

Nyhan was careful not to say that this paper represents a total exoneration of YouTube. The platform hasn’t stopped letting its subscription feature drive traffic to extremists. It also continues to allow users to publish extremist videos. And learning that only a tiny percentage of users stumble across extremist content isn’t the same as learning that no one does; a tiny percentage of a gargantuan user base still represents a large number of people.

This speaks to the broader problem with last month’s new Facebook research as well: Americans want to understand why the country is so dramatically polarized, and people have seen the huge changes in our technology use and information consumption in the years when that polarization became most obvious. But the web changes every day. Things that YouTube no longer wants to host could still find huge audiences, instead, on platforms such as Rumble; most young people now use TikTok, a platform that barely existed when we started talking about the effects of social media. As soon as we start to unravel one mystery about how the internet affects us, another one takes its place.