The studies built to catch a YouTube rabbit hole mostly came up empty

The bubble and the Great Radicalizer

In 2011, the writer and activist Eli Pariser noticed that Facebook had stopped showing him his conservative friends' posts. He hadn't unfollowed anyone. "Facebook was looking at which links I clicked on, and it was noticing that, actually, I was clicking more on my liberal friends' links than on my conservative friends' links. And without consulting me about it, it had edited them out."¹ He named what he'd found the filter bubble: "your own personal, unique universe of information that you live in online."¹ The same tailoring ran underneath the rest of what he used. Google drew on "57 signals" to shape results. Two friends who typed the same search term got different worlds back: one saw the Egyptian protests filling his first page, the other saw nothing about them.¹ Nobody had asked for this. The danger, in Pariser's argument, was an editor nobody could see, narrowing what a person believed everyone else was seeing too.

Facebook's feed and Google's results page are both output by a recommender system: software deciding, out of everything available, what one person sees next. YouTube's version is tuned by engagement optimization, against how long a person keeps watching, not an editor's judgment. A feed built that way reliably rewards attention: it amplifies whatever holds it, calm or extreme. The same ranking shaped Myanmar's feed.

Seven years after Pariser's talk, the technology writer Zeynep Tufekci gave the same worry a sharper edge and a name that stuck: "the Great Radicalizer."² Watching videos of Donald Trump's 2016 rallies, she found YouTube's autoplay queue drifting toward white supremacy and Holocaust denial. A fresh account built on nothing but Clinton and Sanders videos drifted toward left-wing 9/11 conspiracy content.² "It seems as if you are never 'hard core' enough for YouTube's recommendation algorithm," she wrote.² Her explanation was economic: YouTube profits from watch-time, and the algorithm "concluded that people are drawn to content that is more extreme" than what they started with, so escalation pays.² Given YouTube's billion or so users, she wrote, "YouTube leads viewers down a rabbit hole of extremism, while Google racks up the ad sales."²

What four years of real YouTube viewing showed

Tufekci's own case leaned on evidence she called thin: "good data is hard to come by," she wrote.² Her examples came from outside audits, not her own research. Guillaume Chaslot's account-tracking tool, and a Wall Street Journal investigation using the same tool, found that a walk started on either a Clinton or a Trump video was "many times more likely" to end on a pro-Trump one.² Jonathan Albright's search on "crisis actor" found autoplay carrying him into a network of roughly 9,000 conspiracy videos.² All three are the same kind of test: a fresh account with no history, walked through the recommendations to see where it lands. That is not what the algorithm does to a real audience carrying its own history in, and that difference is what the studies that followed measured directly.

The largest of those studies traced real viewing. A team led by the network scientist Homa Hosseinmardi followed 309,813 US adults in a Nielsen panel, across 21,385,962 video views from January 2016 through December 2019.³ Channels were sorted into six political categories; news made up about 11% of what the panel watched.³ Far-right channels barely moved, from 0.17% of watch-time in 2016 to 0.30% in 2019. "Anti-woke" channels, the fastest-growing category, went from 0.31% to 1.02%.³ Sessions of consecutive videos showed no drift toward extremes either.³ Far-right views mostly arrived from outside YouTube's own suggestions: 41.29% followed an external link, 35.8% another YouTube video, the rest from the homepage, a channel page, or search.³ The authors are explicit about the limit: "our analysis ... applies only at the population level. It says little about ... the possible radicalization of individuals or small groups of individuals."³

A second team, led by the political scientists Annie Chen and Brendan Nyhan, tracked actual behavior instead of a panel's viewing history.⁴ They recruited 1,181 people, oversampled for high resentment and heavy YouTube use. A browser extension logged their real viewing and the recommendations YouTube actually showed them, matched to a 2018 attitude survey, in 2020, a year after YouTube overhauled its recommender.⁴ Alternative-media channels made up 3% of logged views; extremist channels, 0.5%; mainstream news, 3.6%; everything else, 93%.⁴ Most of that sliver came from existing subscribers: 60.8% of alternative-channel views, 54.7% of extremist-channel views.⁴ The team applied a strict, four-part definition meant to catch someone with no prior interest walked step by step into extremism. By that definition, rabbit-hole events hit 0.01% of all video visits, among 3.0% of participants, a figure the authors call a lower bound.⁴ "YouTube's algorithms were not sending people down 'rabbit holes' during our observation window in 2020," they wrote, "possibly due to changes that the company made to its recommender system in 2019."⁴

Whose choices actually shape the feed

Pariser's half of the argument got its own test four years after his talk. A Facebook team led by the data scientists Eytan Bakshy, Solomon Messing, and Lada Adamic examined 10.1 million US users who had publicly listed a political affiliation. They tracked 7 million distinct news links their friends shared between July 2014 and January 2015.⁵ They split exposure to this cross-cutting content, meaning news from the other side, into what friends shared, what the feed surfaced, and what users clicked.⁵ Friend networks weren't sealed to begin with: over 20% of an ideology-reporting user's friends listed the opposing party.⁵ Ranking narrowed it 5% for conservatives and 8% for liberals; clicking narrowed it further for conservatives, 17%, though for liberals the two effects were close, 6% versus 8%.⁵ "Compared with algorithmic ranking, individuals' choices played a stronger role in limiting exposure to cross-cutting content," the authors concluded.⁵ The sample skews younger, more educated, and more female than the country, and critics contest measuring ranking against choice rather than a neutral baseline.⁵

The clearest test of whether the ranked feed itself moves anyone's politics came five years later, around the 2020 election. A team led by the political scientist Andrew Guess randomized consenting Facebook and Instagram users onto a chronological feed, no ranking, for roughly three months, against users left on the default algorithm.⁶ Enrollment ran to about 23,000 users per study, per Science's own news coverage rather than the paywalled study text.⁷ The chronological feed changed plenty: less time on the platform, more political and untrustworthy content, and, on Facebook, less uncivil content and more from moderate and ideologically mixed friends.⁶ None of it moved what the study measured: "the chronological feed did not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes during the 3-month study period."⁶

That finding carries its own asterisk. UMass Amherst researchers Chhandak Bagchi and Przemyslaw Grabowicz later found the window wasn't ordinary: "Beginning around the start of November 2020, Meta introduced 63 'break glass' changes to Facebook's news feed, ... expressly designed to diminish the visibility of untrustworthy news surrounding the 2020 U.S. presidential election."⁸ A null measured inside a feed the company had already hand-tuned is narrower than "the ranked feed, as it normally runs, doesn't move anyone's politics."

The gap on both sides

The argument didn't stay confined to a talk and an op-ed. In 2017, Roger McNamee, an early Facebook and Google investor, made the same case from inside the industry. "Facebook, Google and others compete for each consumer's attention," he wrote, "reinforcing biases and reducing the diversity of ideas to which each is exposed."⁹ Same mechanism: attention as the business model, narrowing as the side effect.

Set the studies beside what each can actually reach.

Two of the three concede the individual case outright. Hosseinmardi's team says its method "applies only at the population level" and "says little about ... the possible radicalization of individuals ..."³ Chen's team calls its rabbit-hole rate a lower bound.⁴ The minority that does watch extremist channels, carrying resentments it already held, is real in both datasets. The studies rule out that the algorithm manufactures that minority. They don't rule out that it exists.

Neither side gets to claim these tests as a clean win. "The algorithm radicalizes ordinary users" asks population studies to show something they weren't built to see: one person, over years, on the YouTube that existed before 2019. "Recommenders are harmless" asks a three-month experiment, run inside a feed Meta had already altered, to stand for an ordinary year. Both claims spend confidence the studies did not bank.