Analyzing Right-wing YouTube Channels: Hate, Violence and Discrimination

Motivation

• The study focuses on right-wing YouTube videos and comments because the platform hosts varied right-wing voices and recommendations can lead users to highly partisan channels.The cited newspaper investigation reported this recommendation pattern even for users who had not shown interest in such content.Source: §1 · p. 1

Research questions

• The study asks whether hateful vocabulary, violent content, and discriminatory bias differ in right-wing channels, and whether commenters express hate and discrimination more than video hosts.It also frames the relationship between video content and audience reactions as important to understanding this content.Source: §1 · pp. 1–2

Approach

• The authors propose combining three established procedures into an open-source method for analyzing hate speech and discriminatory bias across text types.They describe the layers as lexicon, topics, and implicit biases, applied to videos and comments and compared with baseline channels.Source: §1 · pp. 1–2

2.1 Dataset

• 3,731 videos and 5,071,728 comments were collected from Alex Jones’ channel and 12 channels he supported, selected after using InfoWars as a seed.Collection used the YouTube Data API from September 28 to October 12, 2017, and included captions and comments, subject to an approximately 500-video-per-channel API limit.Source: §2.1 · p. 2
• 3,942 videos and 12,519,590 comments from ten popular news-and-politics channels formed the baseline, which was not intended to represent neutral user behavior.Channels were selected by subscriber popularity and required mainly English content with non-hard-coded captions available for most videos.Source: §2.1 · p. 2

2.2 Textual preprocessing

• After HTML tags and URLs were removed, language identification retained texts with at least 0.8 probability of being English, yielding 3,278 right-wing videos and 3,581 baseline videos.The retained corpus also contained 4,348,986 right-wing comments and 9,522,597 baseline comments; each video was represented by separate caption and comment documents.Source: §2.2 · p. 2

Three-layered approach

• The three-layered approach evaluates words, topics, and implicit biases to investigate hate, violence, and discriminatory bias.The analyses compare captions with comments and right-wing channels with baseline channels.Source: §3 · p. 3

Lexical analysis

• Lexical analysis compares semantic fields in captions and comments, including fields related to hate, violence, and discrimination, across right-wing and baseline channels.Source: §3 · p. 3

Topic analysis

• Topic analysis contrasts the topics in captions and comments across right-wing and baseline channels.Source: §3 · p. 3

Implicit-bias analysis

• Implicit-bias analysis uses vector spaces that place words sharing contexts near one another to compare biases across sources and channel groups.Source: §3 · p. 3

3.1 Lexical analysis

• The analysis models channel vocabulary through semantic fields, lemmatizing words and normalizing category counts before averaging them across each channel’s videos.The selected Empath categories cover negative and positive matters.Source: §3.1 · p. 3, §3.1.1 · p. 3

3.1.1 Methodology.

• Captions contain more nervousness, rage, and violence terms than comments, while comments contain more swearing; baseline comments have a 6.8% median for hate versus 3.5% in captions.Right-wing captions and comments show no significant difference for hate.Source: §3.1.2 · p. 3
• Right-wing channels have higher fractions of words in categories such as disgust, kill, and terrorism, while baseline channels have higher fractions of joy and optimism.Disgust and swearing terms show no statistical difference between channel types.Source: §3.1.2 · p. 4
• Caption-comment semantic-field similarity varies widely within channels, from near zero in some videos to highly similar in others.The Alex Jones Channel has a median similarity of 0.9, compared with 0.5 for Jon Bowne Reports; baseline-channel distributions have lower variance than right-wing distributions.Source: §3.1.2 · p. 4
• More popular channels generally have higher caption-comment similarity, a pattern the authors suggest may help explain the higher, more consistent baseline values.All baseline channels had at least 1,700,000 subscribers at data collection.Source: §3.1.2 · p. 4
• For comments, right-wing-channel similarity positively correlates with aggression, hate, and violence categories and negatively with positive emotion; baseline comments show a weaker positive correlation with violence.Baseline comments also show a negative correlation with politeness; for captions, significant positive correlations for hate and negative emotion appear only in baseline channels.Source: §3.1.2 · pp. 4–5

3.2.1 Methodology.

• The LDA output lists each document’s top two topics and their 20 highest-ranked words, but topic words can be weakly cohesive or ambiguous, so interpretation is qualitative.The analysis removes punctuation, extra whitespace, and stop words, then lowercases and tokenizes the text; LDA uses up to 2,000 randomly selected tokens per document and returns 300 topics.Source: §3.2.2 · p. 5, §3.2.1 · p. 5

3.2.2 Results.

• Right-wing captions’ top-ranked topics include words related to war, terrorism, espionage, and information war.Examples include “nato,” “torture,” “bombing,” “assange,” and “wikileaks.”Source: §3.2.2 · p. 5
• Right-wing comments’ top-ranked topics include terms possibly related to biological and chemical warfare; “palestinian” may reflect response to “israeli” in captions.Source: §3.2.2 · p. 5
• Right-wing captions and comments generally have more specific topics than baseline channels, whose topics cover a wider range of subjects.Baseline topics include celebrities, TV shows, general news, internet celebrities, and games, with some political interest in “antifa” and “feminists.”Source: §3.2.2 · p. 5

3.3.1 Methodology.

• The three WEATs compare words for immigrants, LGBT people, and Muslims against dominant-group words, using negative and positive attribute sets.The authors measure word–attribute associations and compare associations across the two target classes.Source: §3.3.1 · p. 6
• For each channel, the authors trained separate caption and comment word2vec models, starting from a 600-dimensional Wikipedia model.Source: §3.3.1 · p. 6

3.3.2 Results.

• Figure 4 reports Cohen’s d for biases across channels and sources, showing only p-values below 0.1; Figure 5 aggregates significant biases by channel type and source.Figure 4 marks p-values from 0.05 to 0.1 in orange and below 0.05 in green, with Wikipedia as a dashed reference; signed numbers show comment-minus-caption bias differences.Source: §3.3.2 · p. 7
• Against Wikipedia, Muslim bias was almost always amplified, especially in captions; LGBT bias was weakened in most channels, while immigrant-bias values stayed close to the reference.Wikipedia itself showed relatively high bias, consistent with prior studies.Source: §3.3.2 · p. 7
• In 75% of statistically significant right-wing cases, immigrant and Muslim caption biases exceeded comment biases; for LGBT people, right-wing comments held higher discriminatory bias.Source: §3.3.2 · p. 7
• 1.7 versus 1.5: median Muslim-caption bias was higher for right-wing than baseline channels, but many baseline cases remained inconclusive because fewer showed statistically significant bias.For other topics, differences between channel types were not very pronounced.Source: §3.3.2 · p. 7

• Right-wing channels use more aggression, killing, rage, and violence terms and discuss war and terrorism more, while baseline channels use more joy and optimism terms.Source: §3.4 · p. 8
• Regardless of channel type, YouTube texts amplify discriminatory bias against Muslims and weaken associations of LGBT people with immorality, promiscuity, and sin relative to Wikipedia.Source: §3.4 · p. 8
• The authors conclude that hateful vocabulary and violent content are more accentuated in right-wing channels, with greater anti-Muslim bias in right-wing videos.Source: §3.4 · p. 8
• Comments contain more disgust, hate, and swearing terms, whereas captions contain more aggression, rage, and violence terms.Source: §3.4 · p. 8
• In 75% of right-wing channels, comments show less bias against immigrants and Muslims than captions, but greater LGBT bias; commentators are generally more exacerbated than hosts, with exceptions.Source: §3.4 · p. 8

On hate, violence and bias on the Web

• Hate identification remains an open question: prior work surveys NLP approaches and characterizes hateful users, finding them more negative and profane but counterintuitively less likely to use hate-, terrorism-, violence-, and anger-related words.Source: §4 · p. 9
• YouTube studies have used data mining and social-network analysis to find hate videos, users, and hidden communities, and focused crawlers to mine hate and extremism.Source: §4 · p. 9
• Case studies compare Gamergaters with baseline Twitter users, finding more negative sentiment and hate, while another examines audience participation in an armed-conflict narrative [9].Source: §4 · p. 9

On comment behavior on the Web

• Comment research explores motivations, anonymity, malicious-user detection, and political participation; one classifier for malicious commenters reached almost 81% accuracy.Source: §4 · p. 9
• YouTube video research links popularity with interaction, describing user engagement as ranging from exposure and recommendation to interactivity.Source: §4 · p. 9

5 CONCLUSIONS AND FUTURE WORK

• Right-wing channels discuss war and terrorism more and use more negative word categories, while baseline channels cover more general topics and use more positive words.Source: §5 · p. 9
• Implicit-bias differences for baseline channels are not strong, while right-wing captions show greater bias against immigrants and Muslims and comments show greater bias against LGBT people.Source: §5 · p. 9
• The three established procedures use open-source tools; WEAT complements lexical and topic analyses by measuring implicit associations that word-based measures cannot identify.Source: §5 · p. 9
• Lexical analysis counts words without handling negation, and the study observes data at a single point in time rather than diachronically.Source: §5 · p. 9
• Future temporal analysis could examine whether video behavior stimulates comment behavior and could split aggregated comments into conversation sessions to improve LDA results.Source: §5 · pp. 9–10