Twitch as a privileged locus to analyze young people's attitudes in the climate change debate: a quantitative analysis

The study investigates how Twitch.tv—an emergent, youth-dominated live-streaming platform—hosts discourse on climate change, a topic often examined on Twitter, Facebook, and Reddit. Framed by the COVID-19-accelerated shift to streaming and real-time social interaction, the authors ask whether Twitch’s architecture and commercial aims shape climate topics, how content and interactions unfold in this environment, and whether patterns typical on other platforms (e.g., polarization, echo chambers) are present. The work positions Twitch as a real-time, participatory medium combining live video and synchronous chat, potentially serving as a locus for information, debate, and science communication among young audiences.

Prior research on climate discourse has focused mainly on Twitter, Facebook, Reddit, and YouTube, examining polarization, misinformation, language, and the role of bots. Twitter studies highlight polarization and echo chambers; Reddit and blog comment analyses examine deliberation and event impacts; YouTube work often analyzes comment dominance by elites and anti-activist sentiment, with few studies analyzing video content directly. Twitch research has largely centered on gaming communities, mental health, and platform dynamics, though recent work shows its potential for political activism, affective spamming, and connective democracy. The literature reveals a gap in analyzing Twitch as a site for climate debate and a methodological gap in scalable analysis of video content beyond user comments.

Data were collected on September 9, 2022, using keywords “climate change” and “global warming” on Twitch. From 184 initial entries, 78 streams were retained after filtering duplicates (2), unavailable (29), unrelated (63), and low-impact videos (<5 views; 12). Streams spanned June–September 2022 and involved 56 unique streamers (73.1% male, 26.9% female), mostly broadcasting alone (89.1%). Content categories: games (39.3%), news/politics (26.8%), science/education (17.9%), art/music (10.7%), economy (5.4%). Professionalization: partners 21.4%, affiliates 53.6%, standard members 25%. Data acquisition used the Python Twitch API (channel and stream metadata) and Apify for titles. Audio was extracted and transcribed via Google Cloud Speech API; associated chat logs were downloaded (59 streams had active chat). Workflow: channel data and followers, audio transcripts, chat JSON, and stream metadata were processed; analyses included network analysis, topic/context analysis, media bias/credibility, sentiment, and interaction metrics. Network analysis: Followers of each streamer were retrieved to build an undirected network (nodes=streamers; edges=shared followers). Four isolated nodes were removed for visualization and to avoid fragmentation (final nodes=52; edges=557). Edges were unweighted with a threshold of at least 1 shared follower. Community detection used Louvain’s algorithm (modularity optimization; resolution tuned for meaningful communities). Assortativity coefficient r (Newman) measured mixing across detected communities. Content analysis: Transcripts were cleaned, tokenized, lemmatized, stemmed; stop words, very short, and extremely rare/frequent terms were removed. N-grams (bi/tri-grams) surfaced key phrases (e.g., “global warming,” “Paris Agreement”). Topic modeling employed MALLET LDA (Python wrapper). Coherence Cv guided topic number selection (tested 1–25; optimal 11 topics, Cv=0.417), with lambda fine-tuning (λ=0.2, pyLDAvis). Sampling iterations=1000; hyperparameter optimization interval=10; perplexity P=378.5. Ideological bias and credibility: Named entities (media sources) from transcripts and chat were extracted via Stanford NER, then matched to Media Bias/Fact Check (MBFC) to classify credibility (trusted vs. questionable) and political lean (left/center/right). External URLs from chats were also classified. Context analysis: A Context Index (CI) quantified linguistic formality per stream on a -1 (very informal) to 1 (very formal) scale, built from sentence length, formal language, swearing, and stream category (weighted indicators; preprocessing included sentence tokenization and contraction expansion). CI categories: Very Informal (≤ -0.5), Informal (>-0.5 to ≤ -0.1), Neutral (>-0.1 to <0.1), Formal (≥0.1 to <0.5), Very Formal (≥0.5). View impact categories: Low (1–10), Moderate (11–100), High (101–1000), Very High (>1000 views). Interactions and sentiment: Chat JSONs (59 streams) were cleaned of special characters, URLs, and bot ads; interaction intensity measured as total messages and unique chatters. Sentiment used VADER (lexicon/rule-based; scores -1 to 1), producing positive/neutral/negative proportions per stream. Combined NER + VADER associated sentiments with top-10 named entities.

Network/community: The follower network comprised 52 nodes (after removing 4 isolated from 56) and 557 edges. Three communities emerged: Talk & Play (TP; n≈21), Science & Education (SE; n≈16), and Podcast & Politics (PP; n≈15). Modularity was moderate (0.38), indicating inter-community connectivity. Proximity matrix showed many cross-community ties; SE and PP shared more ties with each other (86) than within their own categories (59–76). Assortativity coefficient r=-0.28 indicated disassortative mixing (non-homophilic), suggesting a non-polarized structure. Top streamers were highly connected both internally and externally (e.g., PP leaders connected to 93% of internal and 80% of external nodes). Interactions: Of 59 streams with chat, nearly 20 had very low impact (<100 comments; <25 unique chatters). Moderate impact was common (32 streams with ~50–100 users and 300–900 messages). Six streams were high impact; one exceeded 400 unique users and 20,000 comments. Recurrent viewers comprised 16–21% of audiences. Sentiment and entities: Sentiment in chats was predominantly positive and neutral, with negative usually <25%; some streams discussing contentious topics showed elevated negativity. Top entities (mentions): Trump (35) and Biden (19) had overall positive sentiment; Congress was prominent among positive mentions, GOP among negative. AOC’s Green New Deal and CCP elicited mixed sentiments. Topics (LDA): Eleven topics captured discourse. The largest was Causes & Consequences of climate change (T1; 36% of tokens), followed by Talk & Play climate discussion (T2; 11%). US politics topics were substantial (e.g., T3 and T4 at ~10% each; T9 at 4%). Skeptical/denial content had low prevalence (3%). Climate science communication (T6) accounted for >6% of tokens. Topic proximity indicated clusters spanning climate, US politics, entertainment, economy, and skepticism. Sources/credibility and bias (MBFC): For 31 streams, over two-thirds (21) of references were to trusted sources; 10 included some questionable sources; 3 were 100% questionable. Ideological bias skewed left/center; right-wing sources appeared in 2 streams. Fourteen streams included references spanning at least two ideological positions; in nine, audiences contributed cross-ideological sources via chat. Context and impact: CI distribution showed many formal/very formal streams (44.9%) alongside informal/very informal (39.7%), with the remainder neutral. Viewing impact was low-to-moderate for ~78% (1–100 views), with a small number of streamers accounting for the most-viewed streams. Overall, chats exhibited positive engagement and manageable sizes conducive to meaningful interaction.

Findings indicate Twitch functions as a real-time ‘third place’ for youth to socialize and deliberate, with climate discourse occurring not only in formal formats (interviews, debates) but also embedded in gameplay (Talk & Play). The non-polarized network structure (r=-0.28; cross-community ties) contrasts with echo chambers typical of Twitter/Facebook studies and suggests shared audiences across content types can reduce homophily. Cross-cutting exposure is present, though it may not always yield deliberation and can devolve into adversarial exchange. The dominance of left/center sources aligns with known environmental attitudes among younger demographics. Despite a generally low presence of questionable sources, a small number of influential streamers referencing questionable media could disproportionately shape narratives given their large audiences and largely positive/neutral sentiment climates. Communities appear cohesive, with moderate chat sizes and recurrent viewership supporting engagement. The presence of both formal and informal formats suggests Twitch’s maturation as a news and science communication venue for climate topics, while raising considerations about traditional media’s growing role and the need to sustain high-quality, independent content.

Twitch hosts an active climate change discourse among young audiences, characterized by cross-community connectivity, generally positive/neutral sentiment, and relatively low prevalence of misinformation compared to findings on other platforms. The study advances research by: (1) focusing on an understudied platform; (2) introducing an objective, scalable methodology combining ASR and NLP to analyze audio content and chats; and (3) centering a youth audience. Results suggest platform architecture and commercial aims do not strictly determine circulating topics; public debate percolates through available technological channels. Policy and practice should recognize Twitch as a venue for science and policy communication, while monitoring the outsized impact of a few influential channels. Future work should expand samples over time, incorporate video/visual context analysis, and track longitudinal shifts in topics, credibility, and network structure as the platform evolves.

Key limitations include: (1) absence of video/visual analysis, limiting contextual cues beyond audio; (2) temporal validity challenges due to Twitch’s ephemerality and changing ecosystem; (3) small sample size (78 streams; 56 streamers; only 59 with chat logs), constrained by limited availability windows (14–60 days) and scraping dates; and (4) potential selection effects from filtering low-view streams and unavailable content. These factors limit generalizability and call for periodic updates and larger, longitudinal datasets.