The idea that YouTube’s algorithm pulls viewers into political “filter bubbles” and “rabbit holes” is widely held. A set of large experiments found little sign of it, at least in the short term.
When researchers tilted real YouTube recommendations toward one side, people did watch more videos from that side. But their opinions on the issues barely changed.
The study was carried out by Naijia Liu, Xinlan Emily Hu, Yasemin Savas, Matthew A. Baum, Adam J. Berinsky, Allison J. B. Chaney, Christopher Lucas, Rei Mariman, Justin de Benedictis-Kessner, Andrew M. Guess, Dean Knox and Brandon M. Stewart, from institutions including Harvard, Princeton, MIT, Duke and the University of Pennsylvania. It first circulated as a Harvard working paper in September 2023 and was published in PNAS on February 18, 2025, as “Short-term exposure to filter-bubble recommendation systems has limited polarization effects: Naturalistic experiments on YouTube.”
A YouTube clone with a dial
The team built its own video platform that looked and worked like YouTube, using real videos and recommendations scraped from YouTube’s own algorithm. Participants watched a starting video on gun control or the minimum wage, then picked their next video from four recommendations, up to five videos in total. Along the way they could like, dislike and save videos.
The researchers then quietly changed the mix. Some participants got a balanced set of recommendations, two from each side. Others got a slanted set, three of four matching the ideological direction of the video they had just watched.
Three filter bubble studies, each with roughly 2,500 participants recruited through MTurk or YouGov, measured attitudes a week before and right after the session. A fourth study, with 1,032 people, tested the rabbit hole idea directly by showing fixed sequences of minimum wage videos that either stayed at the same level of extremity or grew steadily more extreme.
Viewing changed, views did not
The slanted algorithm did shape what people watched. Averaged across studies, liberals given slanted recommendations chose a 6 percentage point larger share of liberal videos than liberals given balanced ones. For conservatives, the liberal share fell by 12 points.
Attitudes hardly budged. Among people with strong views, only one of three studies showed a significant shift: conservatives in the YouGov study became slightly more conservative, by 0.03 on a 0 to 1 policy scale. Moderates showed a similarly small effect in one study. Measures of media trust and hostility toward the other party were essentially unchanged.
The rabbit hole test found nothing either. Watching increasingly extreme sequences had no effect compared with watching sequences of constant extremity. What did shift moderates a little was simply which side’s starting video they were shown.
Because the estimates were tight, the authors argue they can rule out anything beyond quite modest short-term effects. They conclude that “the burden of proof for claims about algorithm-induced polarization has shifted.”
What the experiments cannot rule out
The authors are explicit about the limits. Each session lasted 15 to 30 minutes, so the study cannot speak to months or years of viewing, or to small, vulnerable groups of users. It covers only two policy issues, and only YouTube’s algorithm at the time, which may have changed after earlier criticism.
Almost 90% of participants in the fourth study said they learned something new from the videos, but the type of algorithm did not change how much they reported learning.
Study Details:
- Title: Short-term exposure to filter-bubble recommendation systems has limited polarization effects: Naturalistic experiments on YouTube
- Authors: Naijia Liu, Xinlan Emily Hu, Yasemin Savas, Matthew A. Baum, Adam J. Berinsky, Allison J. B. Chaney, Christopher Lucas, Rei Mariman, Justin de Benedictis-Kessner, Andrew M. Guess, Dean Knox, Brandon M. Stewart
- Journal: Proceedings of the National Academy of Sciences (PNAS)
- Publication Date: February 18, 2025
- DOI: 10.1073/pnas.2318127122