Subscriptions and external links help drive resentful users to alternative and extremist YouTube channels

Contributions

• First direct empirical measurement of how people reach alternative and extremist YouTube videos using real behavioral data (browser extension) paired with survey responses.
• Addresses limitation of prior work by accounting for channel subscriptions—a major predictor of what recommendations users see—and distinguishing internal platform pathways from external referrals.
• Defines and operationalizes "rabbit hole" events with precise criteria; finds them extremely rare, contradicting anecdotal claims of widespread algorithmic radicalization.
• Documents that hostile sexism and racial resentment strongly correlate with consumption of alternative and extremist content; YouTube remains a key distribution channel for this audience.

Method

Study design: Paired behavioral tracking (browser extension) with representative US survey data. Participants recruited from YouGov, oversampling for (1) high gender/racial resentment and (2) frequent YouTube users (n=1,181). Study period: July 21 – December 31, 2020 (after YouTube's 2019 algorithm changes).

Data collection: Browser extension passively logged page views (with URLs and timestamps) and captured HTML snapshots when users viewed YouTube videos, allowing the researchers to extract recommendations and infer subscription status. Subscribe button state (active or inactive) was parsed from each snapshot.

Channel classification: Researchers worked with experts to classify YouTube channels into four types using a curated taxonomy:

• Alternative channels (322 total): discuss controversial topics through a lens attempting to "legitimize discredited views" as marginalized perspectives. Examples: Steven Crowder, Tim Pool, Laura Loomer, Candace Owens, Joe Rogan.
• Extremist channels (290 total): labeled as white identitarian, white supremacist, alt-right, extremist/hateful by sources including the Anti-Defamation League, Center on Extremism, and Stormfront researchers.
• Mainstream media channels (127): meet three criteria—credible NewsGuard score >60, national distribution/cable/print circulation or prestigious journalism award, and US-based.
• Other: channels not fitting the above three categories.

Outcome measures:

• Exposure levels (percentage viewing each channel type).
• Watch time and view counts from alternative/extremist channels, disaggregated by subscription status.
• Concentration of viewership (what fraction of users account for X% of total viewing).
• Correlates: hostile sexism, racial resentment, negative feeling toward Jews, demographics (age, gender, race, education).
• "Rabbit hole" events: strictly defined as recommendations that move users to more extreme channel types and where the user is not already subscribed; very few meet all four operational criteria.
• Internal vs. external referrers: whether users reached videos via YouTube recommendations, searches, channel pages, or external links from other platforms.

Results

Exposure is concentrated among a small, attitudinally distinct audience:

• 91% of participants used YouTube overall; only 15% viewed any alternative channel video, 6% viewed extremist channel videos (vs. 44% for mainstream media).
• Videos from alternative and extremist channels account for only 3.0% and 0.5% of all videos watched (mainstream media: 3.6%).

Subscriptions dominate the viewing pattern:

• 60.8% of views for alternative channel videos come from subscribers to that channel; 54.7% for extremist channels (vs. 38% for mainstream media).
• When including all users subscribed to at least one channel of each type, proportions rise to 92.9% (alternative) and 84.7% (extremist).
• Mean weekly watch time for alternative channel videos: 26 minutes overall (62 min for subscribers, 0.2 min for non-subscribers); extremist: 8 min (15 min for subscribers, 0.04 min for non-subscribers).
• Viewership concentration is severe: 1.7% of users account for 80% of time spent on alternative channel videos; 0.6% for extremist.

"Rabbit holes" are extremely rare:

• Among the study period's video visits, only 0.17% followed a recommendation to an alternative/extremist video (7.3% of participants ever saw such a recommendation).
• Only 0.07% of all visits represent a recommendation moving users to a more extreme channel type (5.4% of participants).
• Probability of recommendation to such content among non-subscribers: negligible (0.02% and 0.01% of visits).
• Total "rabbit hole" events (all four criteria met): 0.01% of visits, observed in only 3.0% of participants. Specifically to extremist channels: 0.002%, only 1.0% of participants.

Recommendation patterns by channel type:

• Alternative/extremist recommendations account for 1.7% and 0.4% of all recommendations shown, respectively.
• When watching alternative channel videos, 47.9% of recommendations point to other alternative channels; 41.1% for extremist videos point to extremist/alternative. When watching mainstream/other content, recommendations to alternative/extremist are rare (<1%).
• Among subscribers to alternative channels, 53.1% of their recommendations were to alternative/extremist videos (vs. 50.3% of all recommendations shown).

Referral pathways:

• Participants often reach alternative/extremist videos via external links (from other platforms like Twitter, Reddit, email) rather than YouTube's internal pathways.
• Prior work found this pattern for far-right news on YouTube; this study replicates and extends it.

Attitudinal predictors:

• Hostile sexism is significantly associated with watching alternative and extremist channel videos (even controlling for other demographics). Effect sizes are large: at sexism level 1 (minimum), predicted viewing is 0.4 min/week (alternative) and 0.08 min/week (extremist); at level 5 (maximum), 383 and 51 minutes/week respectively.
• Racial resentment shows a similar pattern in bivariate analyses but is not significant when controlling for hostile sexism.
• Negative feelings toward Jews: no significant association with alternative/extremist channel viewing.
• Demographics: male, non-white, and more-educated participants tend to watch more alternative channel videos. Bachelor's and post-grad education predict higher extremist viewing.

Notes

Strengths:

• Direct behavioral measurement via browser extension avoids ecological validity problems of lab experiments or recall bias of surveys alone.
• Controlled for channel subscriptions, a key factor prior work overlooked. Subscriptions heavily influence what recommendations users see.
• Precise operationalization of "rabbit hole" events (four criteria must all hold), raising the bar for claims about algorithmic radicalization.
• Large, weighted sample designed to approximate US adult population on key demographics; oversampled vulnerable groups.
• Ethical research design with full informed consent, IRB approval, and data housed securely.

Limitations and caveats:

• Study period (July–Dec 2020) is post-2019 algorithm changes; cannot assess radicalization risk before YouTube intervened. Results represent a "lower bound" on rabbit-hole events — the true rate pre-2019 may have been higher.
• Small fraction of potential extremist audience: 1,181 participants is a targeted sample, not a census. Findings may not generalize to users who don't volunteer for studies or who are already heavily radicalized.
• Cannot prove causation from correlation: high resentment correlates with extremist viewing, but causality is unclear (does consumption drive resentment, or vice versa?).
• Browser extension only captures desktop/Chrome/Firefox; mobile and app-based viewing is excluded.
• Subscription status inferred indirectly from subscribe button state at time of viewing; users may subscribe and unsubscribe between videos, and the study does not have full subscription history.
• Researcher judgement required to classify channels as "alternative" or "extremist." Though expert-reviewed, this taxonomy is contextual and may shift over time or differ across jurisdictions.
• No direct measure of eye gaze or active viewing; watch time is inferred from page view timestamps capped at video length.

Significance: This paper provides the strongest empirical evidence to date that YouTube radicalization, if it occurs via recommendations, is rare post-2019. The study validates concerns about platform-hosted extremist content while shifting responsibility from the algorithm to user demand and deliberate external promotion. It reframes the policy debate: the question is not "does the algorithm radicalize?" but "how do we address demand for extremist content among resentful users, and how do we prevent external actors from promoting radicalization pathways?"

The finding that subscriptions and external links matter more than recommendations has major implications for platform design (should subscriptions be moderated differently?) and for research (past reliance on random-walk simulation underestimated the role of deliberate user seeking).