A Claim Repeated So Often It's Treated as Settled
The idea that YouTube's recommendation algorithm systematically pulls viewers down a "rabbit hole" toward increasingly extreme content has become one of the most widely repeated claims about social media's harms - cited in journalism, policy debates, and casual conversation as established fact. The actual body of academic research on the question is considerably more mixed than that popular narrative suggests, and the disagreement between studies isn't a minor technical dispute - it points to genuinely different conclusions about where responsibility for radicalisation actually lies.
The Studies That Support the "Rabbit Hole" Narrative
A systematic review of 23 academic studies on the topic found 14 implicated YouTube's recommendation system in facilitating pathways toward more extreme content, while only 2 found no such link. A widely cited 2020 study by researchers Ledwich and Zaitsev found troubling patterns in early recommendation behaviour, and a 2023 study published in PNAS found YouTube tends to recommend ideologically congenial content to already-partisan users, with recommendations becoming more problematic deeper into a viewing session, an effect the researchers found was more pronounced for right-leaning users specifically.
The Studies That Push Back Hard
A significant and growing body of more recent, methodologically rigorous research complicates this picture considerably. A 2024 study using counterfactual "bot" accounts - designed to isolate the algorithm's specific effect from a user's own choices - found that on average, YouTube's recommender did not radicalise users further, and for some users, actually had a moderating effect, pulling them toward more mainstream content rather than away from it. A 2025 University of Pennsylvania study reached a similar conclusion, finding limited evidence that recommendation algorithms alone drive meaningful shifts in users' political views, and suggesting that a user's own pre-existing interests and active choices - not passive algorithmic push - play the primary role in what gets watched.
Why Researchers Disagree So Much
Part of the disagreement comes down to methodology - studies using "sock puppet" accounts that passively follow every recommendation tend to find stronger algorithmic effects than studies that model more realistic viewing behaviour, where users actively search, subscribe, and click based on their own interests rather than passively accepting whatever is suggested next. There's also a platform-specific dimension: research specifically on TikTok has found more consistent evidence of rapid algorithmic steering toward extreme content than equivalent YouTube research, suggesting the answer may genuinely differ by platform, format, and recommendation design rather than there being one universal answer for "algorithms" as a category.
The More Defensible Conclusion
The most careful current research suggests a more nuanced picture than either "algorithms radicalise people" or "algorithms are blameless." Recommendation systems appear to more reliably amplify and reinforce a viewer's existing interests and leanings than to independently manufacture new extreme views in someone who wasn't already inclined that way - meaning off-platform factors (community, ideology already held, real-world grievances) likely play at least as large a role as the specific mechanics of any single platform's recommendation engine, even as the platform can still meaningfully shape how much of that content someone encounters and how quickly.
Why Getting This Right Actually Matters
This isn't just an academic distinction. If algorithms are the primary driver, the fix is largely technical - adjusting recommendation systems. If a user's own existing interests and offline context are the primary driver, technical fixes alone won't be enough, and addressing radicalisation requires looking well beyond any single platform's recommendation code. The honest, current answer is that both algorithmic amplification and individual predisposition likely play a role, in proportions that current research hasn't definitively settled - which is a less satisfying answer than either side's confident claim, but a more accurate one.
