Auditing Radicalization Pathways on YouTube

1 INTRODUCTION

• The audit examined more than 72 million comments across 330,925 videos from 349 channels, treating Alt-right content as a proxy for extreme content.It compared the I.D.W., Alt-lite, and Alt-right communities, with media channels included as a growth sanity check.Source: §1 · p. 2
• Users who initially commented only on I.D.W. or Alt-lite content increasingly began commenting on Alt-right content, forming a significant share of Alt-right commenters.The authors interpret their findings as strong evidence of user radicalization, while treating commenters as a proxy for users consuming content.Source: §1 · p. 2
• About half of users who commented on Alt-right channels in 2018 also commented on Alt-lite and I.D.W. channels, and the communities increasingly shared users.Source: §1 · p. 2
• Recommendation simulations found Alt-lite channels easily discoverable from I.D.W. channels and Alt-right channels reachable from the other communities, particularly through channel recommendations.The collected recommendations did not account for personalization.Source: §1 · p. 2
• The communities showed steep growth in activity and engagement relative to media channels, with Alt-right channels reaching nearly one comment per five views in 2018.Source: §1 · p. 2

Contrarian communities

• The I.D.W., Alt-lite, and Alt-right are framed as contrarian communities shaped by 2010s anti-PC culture.The I.D.W. describes academics and media personalities discussing controversial subjects.Source: §2 · p. 2
• The Alt-right is associated with white-supremacist and racial-nationalist ideology, while the Alt-lite distinction is blurry, so the study labels only the most extreme creators Alt-right.The study places borderline I.D.W. cases in the Alt-lite; boundaries are especially unclear for non-core members.Source: §2 · pp. 2–3

Radicalization

• The study defines radicalization as movement toward greater justification of intergroup violence and uses increased Alt-right content consumption as its proxy.It argues this proxy is reasonable because Alt-right rhetoric has been invoked by terrorist-attack perpetrators and promotes ideas such as a white ethnostate; conservative labeling gives the Alt-lite and I.D.W. the benefit of doubt.Source: §2 · p. 3

Auditing the web

• The audit measures user radicalization in an algorithm-influenced content-sharing environment, but analyzes only a time slice of recommendations and does not account for personalization.The authors note that comment histories can offer insight into the past, but cannot readily isolate past algorithmic influence; they say the framework can be replicated over time and expanded to handle personalization.Source: §2 · p. 3

Previous research from/on YouTube

• Prior YouTube research examined recommendation-system technicalities, harmful-content detection, extremist communities, channel topics, and metadata-based recommendation [32].Sureka et al. tracked extremist communities using user connections, subscriptions, and favorites, while Ottoni et al. analyzed topics across 23 channels [32] and O’Callegan et al. found metadata-based recommendations could cluster extreme-right topics.Source: §2 · p. 3

3 DATA COLLECTION

• Channel pools began with representative seeds sourced from,, and, expanded through the first 200 English YouTube search results per keyword, and iteratively added relevant featured and recommended channels.The authors collected two additional hops of featured or recommended channels; these links could come from creators’ featured tabs or YouTube’s recommender system.Source: §3 · p. 3
• For each channel, the researchers collected subscribers and views, video comments and captions, plus 22 rounds of channel and 19 rounds of video recommendations.They used VPNs from seven locations, visited channels in random order, and did not log into accounts, so the collection did not include personalization.Source: §3 · p. 4
• The final labels covered 85 I.D.W., 112 Alt-lite, and 84 Alt-right channels, with 75.57% interannotator agreement (95% CI [67.5, 82.5]).Two experienced authors inspected channels’ popular videos and watched at least five minutes of content; disagreements were discussed until they reached a conclusion.Source: §3 · p. 4
• The researchers collected 68 popular media channels from Media Bias/Fact Check categories as a sanity check for trends among mainstream YouTube channels.Channels were included when a YouTube search for the source name matched on the first results page; some were excluded because they had more than 15,000 videos.Source: §3 · p. 4

Recent rise in activity

• Likes, views, and comments grew exponentially across all communities and media channels, while channel creation in the communities of interest rose later than in media channels.In mid-2015, 66 of 68 media channels were active, compared with fewer than half of Alt-lite, Alt-right, and I.D.W. channels; the authors relate community growth to the 2016 presidential campaign.Source: §4 · p. 5

Engagement

• Alt-right channels had the highest comments per view from 2017, reaching nearly 1 comment per 5 views by 2018, while communities of interest overall showed more engagement than media channels.Media channels had more views per video, but those views were less often converted into likes and comments; Alt-right engagement growth followed that of I.D.W. and Alt-lite, consistent with a proposed fertile-ground narrative.Source: §4 · p. 5

Dormant Alt-right Channels

• By 2013, approximately 30 channels in each of the three communities were active, but Alt-right channels had published a bit more than 20% of their content, versus nearly 40% for media and Alt-lite channels.Source: §4 · p. 5

• These descriptive trends cannot determine whether a radicalization pipeline exists; the following sections examine the relationships between communities through users who commented on them.Source: §4 · p. 5

5 USER INTERSECTION

• The analysis compared user groups with Jaccard similarity and the Overlap Coefficient, which is useful for comparing communities of different sizes.A small subset of a large set can have low Jaccard similarity but an Overlap Coefficient of 1.Source: §5 · p. 6
• Communities of interest had many more highly active commenters than media channels, although Alt-right commenters averaged fewer comments than Alt-lite or I.D.W. commenters.Because the Alt-right community was younger, the paper cautions that this difference does not establish lower engagement.Source: §5 · p. 6
• User retention increased over time in the I.D.W., Alt-lite, and Alt-right, while media Jaccard similarity plateaued from 2014 and its overlap coefficient only recently began to grow.Commenters in the communities of interest seemed to return more often than media commenters.Source: §5 · p. 6
• In 2018, Alt-lite–I.D.W. Jaccard similarity reached almost 30%, while around 50% of Alt-right commenters also commented in the Alt-lite and I.D.W.The Alt-lite–I.D.W. Jaccard similarity exceeded the self-similarity of either community.Source: §5 · p. 6
• Jaccard similarity with media was comparable to similarity with Alt-right, but overlap with media remained lower than self-overlap, especially since 2015.Media similarity grew sharply in 2018, possibly as the communities became more mainstream; the communities also formed a significant portion of the large media-channel user base.Source: §5 · p. 6

User migration

• Tracking is limited to collected channels; cohorts without prior Alt-right comments are followed by year, with exposure defined as 1–2 (light), 3–5 (mild), or 6+ (severe) Alt-right videos.Users tracked for one year are not eligible for selection in upcoming years.Source: §6 · p. 7
• Among the 1,251,674 users who commented only on Alt-lite or I.D.W. videos in 2017, approximately 12% were exposed in 2018—more than 60K users altogether.Source: §6 · p. 7
• In the last three time brackets, under 1% of media-channel users were mildly or severely exposed, versus 3%–4% of Alt-lite or I.D.W. users; light exposure was roughly 4% versus 8%.The passage reports lower exposure rates for media-channel users than for Alt-lite or I.D.W. users.Source: §6 · p. 7
• For users tracked from 2006–2012, exposure was around 15% for both Alt-lite-only and I.D.W.-only commenters; among 2017 cohorts, it was about 12% versus 6%.The Alt-lite/I.D.W. difference grew over time.Source: §6 · p. 7
• In 2018, roughly 40% of Alt-right commenting users across exposure levels could be traced to earlier cohorts who commented only on Alt-lite or I.D.W. videos.The figure measures their share of the total Alt-right commenting audience.Source: §6 · p. 8, Figure 4 · p. 8
• In 2018, 23.3% of lightly exposed users traced back to prior Alt-lite commenters, versus 7.6% to prior I.D.W. commenters; the I.D.W. share was around 3 times lower.Across exposure levels and time points, the share traced to I.D.W.-only commenters was around 3 times lower than the share traced to Alt-lite-only commenters.Source: §6 · p. 8

7 THE RECOMMENDATION ALGORITHM

• The study combines weighted channel recommendations into a graph and runs 10K five-step random walks from subscriber-weighted starting nodes, with separate rules for “Other” in channel and video graphs.The recommendation data are a non-personalized snapshot, limiting conclusions about the recommender’s role in radicalization; video walks exclude “Other” because it receives too many edges, while channel walks stop there.Source: §7 · pp. 8–9
• Alt-lite and I.D.W. channel recommendations stay within-community about 50% of the time and cross between them 14% or 23%; Alt-right channels are recommended only by Alt-lite (3.08%).More than 75% of video recommendations from every community point to untracked “Other” videos; Alt-right videos are not significantly recommended.Source: Table 3 · p. 8, §7 · p. 9
• In channel walks, Alt-right reachability@5 is approximately 4% from Alt-lite and 1.5% from I.D.W.; I.D.W.-started walks find Alt-lite within five steps in 25% of runs.Starting from an I.D.W. channel, there is approximately a 10% chance of being in an Alt-lite channel at the next step.Source: §7 · pp. 9–10, Figure 5 · p. 9
• From media channels, channel-walk reachability@5 is 2.5% for I.D.W. and slightly less than 1% for Alt-lite.In video recommendations, media channels are recommended at around 3% across communities, while Alt-lite and I.D.W. recommend each other roughly 2% of the time.Source: §7 · pp. 9–10
• In video walks, Alt-right reachability@5 from Alt-lite is around 0.05%; I.D.W.-to-Alt-lite reachability@5 is roughly 7%, with untracked videos excluded.From media channels, video-walk reachability@5 is around 4.5% for I.D.W. and 1.5% for Alt-lite; this setup is less realistic because it ignores untracked videos.Source: §7 · pp. 9–10, Figure 5 · p. 9

Discussion

• A significant number of commenting users consistently migrated from milder content toward Alt-right content, which the authors interpret as evidence of ongoing YouTube radicalization.The migration was consistent across years and significant in absolute quantity; Alt-right channels serve as a proxy for extreme content.Source: §8 · p. 10
• The three communities’ activity rose particularly after 2015, while their users commented more assiduously than media-channel users and extreme content had more comments per view.The discussion notes that larger media channels also grew, and that the communities’ growth coincided with the 2015 presidential election.Source: §8 · p. 10
• Commenting users are treated as a radicalization proxy partly because only 5 of 900 sampled comments criticized their associated videos.The discussion also reports a higher proportion of likes for the communities of interest (> 91% mean, > 96% median) than for media channels (85% mean, 93% median).Source: §8 · p. 10
• Recommendations frequently surfaced Alt-lite and I.D.W. content, with Alt-right content reachable from them through channel—but not video—recommendations.The authors caution that they sampled only a tiny fraction of recommendations and did not account for personalization; they nevertheless found a path to extreme content from large media channels.Source: §8 · p. 10

Limitations and future work

• The study measures traces of radicalization but offers little explanation for why pathways exist, motivating research on causes and changes in creators’ and users’ speech over time.The analyses focus on likes, views, comments, user trajectories inferred from comments, and simulated trajectories in recommendation graphs.Source: §8 · p. 10

• Tables 7–8 document collection steps for Alt-right, Alt-lite, and I.D.W. channels, Table 9 lists media channels, and Figure 9 highlights the collected YouTube material.Source: A DATA COLLECTION · p. 12
• Channel searches used category-specific seeds, including I.D.W. figures such as Stephen Hicks, Alt-lite figures such as Brittany Pettibone, and Alt-right terms and figures such as “White Genocide” and Richard Spencer.Source: A DATA COLLECTION · p. 12

B FEATURED VS RECOMMENDED

• Featured channels are chosen by the channel owner, whereas related channels are recommendations made by YouTube.The distinction is illustrated in Fig. 8 for featured channels and Fig. 9(e) for related channels.Source: B FEATURED VS RECOMMENDED · p. 12

C LIKES, VIDEOS, VIEWS, COMMENTS

• Tab. 4 reports yearly counts of likes, views, videos, and commenting users across the three communities.Source: C LIKES, VIDEOS, VIEWS, COMMENTS · p. 12

D USER TRAJECTORIES

• Alt-right users diffuse to the other communities and media channels in similar proportions, with Tables 5 and 6 reporting tracked-user counts and the infected share of Alt-right viewers.Table 6 repeats the trajectory procedure in the reverse direction, checking whether users who commented in Alt-right also commented in the other communities or media channels.Source: D USER TRAJECTORIES · p. 12

• Figures 6 and 7 show the recommendation graphs used for the Section 7 experiment.Source: E RECOMMENDATION GRAPHS · p. 12