Where do cross-cutting discussions happen?: Identifying cross-cutting comments on YouTube videos of political vloggers and mainstream news outlets

The study investigates whether social media, particularly YouTube, fosters echo chambers or facilitates cross-cutting discussions among political groups. Building on debates around selective exposure and algorithmic reinforcement of echo chambers, the authors focus on political vloggers compared with mainstream news outlets. They define "cross-cutting comments" as those opposing the channel's political leaning and pose six research questions: RQ1-1: To what extent do cross-partisan discussions occupy conservative and liberal vloggers' comment threads? RQ1-2: Does the extent of cross-partisan discussions in a comment thread significantly vary by the vlogger's political leaning? RQ2: How well can comments from vlogger videos work as training data to predict the political leanings of comments from mainstream news outlet videos? RQ3-1: Does the proportion of cross-cutting discussions significantly vary between conservative and liberal mainstream news outlet videos? RQ3-2: In what ratio do conservative and liberal comments make up the comment thread of the neutral mainstream news outlet video? RQ4: Is the proportion of cross-cutting comments in mainstream news outlet comment threads significantly higher than that in vlogger comment threads? The study uses manual coding of vlogger comments and machine learning classification of mainstream news outlet comments on a single issue (the Mueller report) to avoid topic heterogeneity and enhance NLP classification reliability.

Echo chambers describe environments where individuals are exposed predominantly to information that reinforces their existing beliefs. Concerns intensified with social media and recommendation algorithms that may curate ideologically homogeneous content. However, empirical evidence is mixed: some work finds no avoidance of opinion-challenging content online and highlights the presence of "cross-cutting discussions"—dialogues among people with differing perspectives. Wu and Resnick (2021) showed asymmetric cross-cutting on YouTube: conservatives comment more on left-leaning videos than liberals do on right-leaning ones, and mainstream outlets host more cross-cutting than independent channels. Prior research also suggests parasocial relationships between YouTube influencers and their audiences, potentially reinforcing alignment with the influencer's views and hindering cross-cutting dialogues. The current study addresses gaps by focusing specifically on political vloggers (distinct from organizational independent media) and examining how media type and outlet leaning are associated with cross-cutting discussions, including the role of a neutral outlet (C-SPAN).

Design: Two-stage analysis. (1) Manual coding of comments on political vlogger videos; (2) Supervised machine learning classification (multiclass) of comments on mainstream news outlet videos using vlogger comments as training data. Topic: Single-issue focus on the Mueller report (March–April 2019) to reduce NLP complexity from topic heterogeneity and improve detection of group-specific linguistic patterns. Data collection windows: Videos and comments between Mueller’s submission (03/22/2019) and AG Barr’s press conference (04/18/2019). Vlogger data: Judgment sampling of top-subscribed political vloggers from Feedspot and Socialblade; inclusion required a Mueller-related upload within one week of 03/22/2019, ≥100,000 views or ≥100 comments, and clear conservative/liberal leaning. Ten videos total (5 conservative, 5 liberal). Comments were collected via YouTube API, replies excluded, and 100 random top-level comments per video sampled, yielding 1,000 comments (500 conservative-leaning vlogger videos, 500 liberal-leaning vlogger videos). Manual coding: Two trained coders used a codebook to label each comment as conservative, liberal, other, or indeterminable. Intercoder reliability: Gwet’s AC1 = .835 (> .800 threshold). Labels were later binarized to cross-cutting vs. non-cross-cutting for chi-square tests. Mainstream outlet data: Seven videos of AG Barr’s press conference (04/18/2019) from Fox News, LiveNOW from FOX (conservative), CNN and MSNBC (liberal), and C-SPAN (neutral). All were raw event footage without commentary, uploaded on the same day. Comments collected via YouTube API; replies excluded; final sample size 4,230 comments. NLP preprocessing and modeling: spaCy used to remove 326 stop words and tokenize. Sentence embeddings via Sentence-BERT (SBERT), model all-mpnet-base-v2. Three supervised classifiers: logistic regression, support vector machine (SVM), and random forest. Multiclass labels: conservative, liberal, neither (combining "other" and "indeterminable"). Evaluation: 200 random mainstream comments were human-coded (same coder as manual stage) and compared against model predictions; metrics computed (accuracy, macro-F1, macro-precision, macro-recall). Baseline accuracy = .333. Models surpassing baseline were used to predict remaining 4,030 comments. Statistical tests: Chi-square tests assessed differences in distributions and cross-cutting proportions by channel leaning; Monte Carlo simulation used for chi-square with low cell counts in random forest results. Comparison across media types: The model with highest accuracy (logistic regression) was used to compare cross-cutting proportions between vlogger and mainstream outlet comments.

Vlogger comments (manual): Conservative vlogger videos: 76.2% conservative (381/500), 3.0% liberal (15/500), 4.0% other (20/500), 16.8% indeterminable (84/500). Liberal vlogger videos: 70.0% liberal (350/500), 10.0% conservative (50/500), 1.0% other (5/500), 19.0% indeterminable (95/500). Cross-cutting proportions: Liberal vlogger threads had 10.0% cross-cutting vs. 3.0% in conservative vlogger threads; difference significant, X²(1, N=1000)=19.02, p<.001. Model performance on mainstream comments (200 evaluated): Logistic regression: accuracy .640, macro-F1 .578, macro-precision .623, macro-recall .568. SVM: accuracy .615, macro-F1 .558, macro-precision .567, macro-recall .567. Random forest: accuracy .610, macro-F1 .439, macro-precision .562, macro-recall .446. Mainstream predictions (full 4,230): Logistic regression (best model): Conservative outlets (N=2,421): 53.0% conservative (1,282), 39.2% liberal (949), 7.9% neither (190). Liberal outlets (N=1,672): 54.4% conservative (909), 42.1% liberal (704), 3.5% neither (59). Neutral outlet (C-SPAN, N=137): 46.7% conservative (64), 45.3% liberal (62), 8.0% neither (11); ratio conservative:liberal ≈ 1.03:1. Overall distribution varies by outlet leaning: X²(4, N=4230)=35.20, p<.001. Cross-cutting proportions: Liberal outlets’ cross-cutting (conservative comments on liberal channels) 54.4% vs. conservative outlets’ cross-cutting (liberal comments on conservative channels) 39.2%; X²(1, N=4093)=91.17, p<.001. SVM results: Conservative outlets: 48.5% conservative (1,175), 39.0% liberal (943), 12.5% neither (303). Liberal outlets: 49.5% conservative (827), 43.7% liberal (730), 6.9% neither (115). Neutral: 45.3% conservative (62), 45.3% liberal (62), 9.5% neither (13); distribution varies: X²(4, N=4230)=37.42, p<.001; cross-cutting difference: X²(1, N=4093)=44.09, p<.001. Random forest results: Conservative outlets: 71.5% conservative (1,730), 27.4% liberal (663), 1.2% neither (28). Liberal outlets: 66.6% conservative (1,113), 32.6% liberal (545), 0.8% neither (14). Neutral: 62.0% conservative (85), 37.2% liberal (51), 0.7% neither (1); distribution varies: X²=17.25, p=.004 (Monte Carlo); cross-cutting difference: X²(1, N=4093)=616.52, p<.001. Media type comparison (using logistic regression): Conservative channels: mainstream cross-cutting 39.2% vs. vlogger 3.0%; X²(1, N=2921)=91.17, p<.001. Liberal channels: mainstream cross-cutting 54.4% vs. vlogger 10.0%; X²(1, N=2172)=243.96, p<.001. Overall: Mainstream news outlet threads have substantially higher cross-cutting proportions than vlogger threads across political leanings.

Findings confirm asymmetric cross-cutting participation on YouTube: more cross-cutting on liberal channels than conservative ones, both for vloggers (manual) and mainstream outlets (computational), aligning with prior work. Media type matters: mainstream outlets display much higher cross-cutting than vlogger channels, suggesting that personal influencer-channel dynamics (including parasocial relationships) may foster more ideologically homogeneous communities and potential echo chambers among vloggers’ audiences. Qualitative observations reveal that some cross-cutting on mainstream outlets involves ridicule or trolling, indicating that sheer cross-cutting volume may not equate to constructive discourse. Neutral outlets (C-SPAN) show the most balanced conservative–liberal mix and comparatively more rational discussion, supporting the potential of politically neutral venues for productive cross-partisan engagement. Context also influenced outcomes: given that the Mueller report’s contemporaneous framing favored conservatives, conservative comments dominated even liberal outlet threads; liberals appeared less engaged, while conservatives celebrated and taunted, explaining counterintuitive distributions. Vlogger threads, by contrast, remained aligned with the vlogger’s leaning, suggesting stability and reduced susceptibility to short-term news context, consistent with community/fandom dynamics on YouTube.

The study demonstrates that cross-cutting discussions on YouTube vary by both channel political leaning and media type. Liberal channels exhibit higher cross-cutting than conservative channels; mainstream news outlets host more cross-cutting than political vloggers. Neutral news outlets (C-SPAN) appear promising as venues for balanced, constructive cross-partisan exchanges. Methodologically, vlogger comments can serve as effective training data for classifying political leanings in mainstream outlet comments via multiclass NLP models, achieving accuracy well above baseline. Qualitative insights caution that cross-cutting volume alone does not ensure productive discourse and that news context critically shapes comment distributions. The work advances understanding of cross-partisan communication in video-based social media and points to future research on recommendation networks, community culture, and the role of neutral outlets.

Generalizability is constrained by the single-issue design (Mueller report), chosen to enhance NLP reliability by reducing topic heterogeneity; results may differ for other issues. Multiclass classifiers had limited accuracy (though above baseline), so findings should be interpreted cautiously. Replies were excluded, potentially omitting interactional nuance. The study did not account for each vlogger channel’s community culture, which may affect cross-cutting dynamics. Neutral outlet sampling included only C-SPAN due to difficulty identifying other truly neutral mainstream channels.