Back to the directoryFind in the timeline
YouTube radicalisation audits and YouTube's borderline content post
Studies disagree on whether YouTube's recommendations push viewers toward extreme content. Ribeiro and colleagues found movement from milder to more extreme channels, while Hosseinmardi and colleagues and Chen and colleagues found small, concentrated audiences and no systematic recommendation effect.
- Where it sits
- Academic psychologyResearch published for other researchers and students.
- Evidence level
- ContestedThe evidence conflicts or is disputed, for example after failed replications.
- Organizations
- YouTubeAn organization is listed where the sources concern it. The text says exactly how.
- Sources inspected
- 5 of 52 with the relevant text, company document or archive item checked, 3 as an abstract or reference article, 0 not verified.
What happened
On 25 January 2019 YouTube said it would begin reducing recommendations of borderline content. An audit by Ribeiro and colleagues, first posted on arXiv in August 2019, covered 349 channels, with recommendations collected from May to July 2019. Hosseinmardi and colleagues (PNAS, 2021) analyzed browsing by a panel of more than 300,000 Americans from 2016 to 2019. Chen and colleagues (Science Advances, 2023) used browser activity from 1,181 US adults in 2020.
The psychological idea
The designs differ. Ribeiro and colleagues grouped channels into four types, processed over 72 million comments and probed more than 2 million recommendations. Hosseinmardi and colleagues used a representative browsing panel, on and off YouTube. Chen and colleagues combined participants' browser activity with 2018 survey measures of racial and gender resentment. YouTube's post is the company's own statement, not an audit.
In practice
Ribeiro and colleagues report that users tend to move from milder to more extreme content and that Alt-right videos were reachable only through channel recommendations. Hosseinmardi and colleagues report mostly mainstream news viewing, a small stable far-right audience and no evidence that recommendations systematically cause far-right engagement. Chen and colleagues report that about 1.7 per cent of participants accounted for 80 per cent of alternative-channel watch time and that subscribers supplied 60.8 per cent of alternative-channel views.
Why it mattered
The studies show how method shapes the answer: channel audits, audience overlap and real browsing histories can point different ways. They are also not simple rivals, because they cover different years and different versions of YouTube. Chen and colleagues describe their own results as consistent with both earlier papers.
Limits and qualifications
Contested. YouTube's post says its change applies to less than one per cent of content and gives no watch-time share, so it is self-report. Chen and colleagues say they cannot assess YouTube's earlier recommendations, that their rabbit-hole counts are lower bounds, and that their data are observational, desktop-only and US-only. Only abstracts of the other two studies were read, and later arXiv revisions of the Ribeiro paper may differ.
Evidence behind the claims
YouTube's post of 25 January 2019 says it will begin reducing recommendations of borderline content, that this applies to less than one per cent of content, and gives no watch-time figure.
Ribeiro and colleagues audited 349 channels in four types, with 330,925 videos, over 72 million comments and more than 2 million recommendations (recommendations collected May to July 2019), and report that users tend to move from milder to more extreme content and that Alt-right videos are reachable only through channel recommendations.
Hosseinmardi and colleagues, using a representative panel of more than 300,000 Americans from January 2016 to December 2019, report news consumption dominated by mainstream sources and a small, stable far-right share of consumers.
The Hosseinmardi abstract states the authors find no evidence that engagement with far-right content is caused by YouTube recommendations systematically, and no clear evidence that anti-woke channels serve as a gateway to the far right.
Chen and colleagues, with 1,181 US adults tracked from 21 July to 31 December 2020, report that 15.4 per cent watched an alternative channel video, 1.7 per cent of participants accounted for 80 per cent of alternative-channel watch time, and subscribers supplied 60.8 per cent of alternative-channel views.
Chen and colleagues say their data cannot evaluate the earlier state of the platform, treat their rabbit-hole counts as a lower bound, make no causal claims, and cover only desktop Chrome and Firefox and US respondents.
Chen and colleagues describe their findings as replicating Hosseinmardi and colleagues on audience size and referral patterns, and their recommendation patterns as substantively similar to random-walk studies including Ribeiro and colleagues, though they do not directly replicate Ribeiro's methods.
People involved
- Manoel Horta Ribeiro. Lead author of the audit preprint
- Homa Hosseinmardi. Lead author of the 2021 PNAS study
- Annie Y. Chen. Lead author of the 2023 Science Advances study
- Brendan Nyhan. Co-author of the 2023 study
In the Psy101 course
- OpenStax Psychology 2e, section 2.3: Analyzing FindingsChapter 2, Psychological Research. Method. The section discusses this topic.The analyzing findings section covers correlation versus causation and experiments, relevant to observational audits.
This is course relevance only. It is not evidence for the historical or corporate claims on this page.See the full course map.
Sources and inspection
5 of 5 sources were opened and read at the level shown. A source marked as not verified supports no claim on this page by itself.
- [1]Auditing Radicalization Pathways on YouTubeRibeiro, M; Ottoni, R; West, R; Almeida, V; Meira, W. arXiv 1908.08313, 2019. Preprint. Primary source.Abstract or publisher record onlyRead the arXiv abstract page and version list (v1 August 2019 to v4 October 2021). The page names no venue; counts are those shown on the current page.
- [2]Examining the consumption of radical content on YouTubeHosseinmardi, H; Ghasemian, A; Clauset, A; Mobius, M; Rothschild, D; Watts, D. Proceedings of the National Academy of Sciences 118 (record on IDEAS/RePEc), 2021. Journal article. Primary source.Abstract or publisher record onlyRead the IDEAS/RePEc abstract and publication details (PNAS 118(32), August 2021).
- [3]Subscriptions and external links help drive resentful users to alternative and extremist YouTube channelsChen, A; Nyhan, B; Reifler, J; Robertson, R; Wilson, C. Science Advances 9(35), PubMed Central copy, 2023. Journal article. Primary source.Relevant text checkedRead the open-access PMC article as returned by the fetch tool: sample, key results, limitations and discussion of prior work. The whole paper was not read.
- [4]Continuing our work to improve recommendations on YouTubeYouTube. YouTube Official Blog, 2019. Company document. Company document.Company document readRead the company blog post as returned by the fetch tool. Company self-report, not an audit.
- [5]Examining the consumption of radical content on YouTube (arXiv version)Hosseinmardi, H; Ghasemian, A; Clauset, A; Mobius, M; Rothschild, D; Watts, D. arXiv 2011.12843, 2020. Preprint. Primary source.Abstract or publisher record onlyRead the arXiv abstract of the same paper, which contains the statements on recommendations and the gateway hypothesis.
Access note: Only abstracts of the Ribeiro and Hosseinmardi papers were read. The Ribeiro venue is not stated on the page opened, and later arXiv revisions may report different counts.
Independent source review: Claude review agent, independent of the research agent, 2026-10-21. Verdict: pass-with-edits.