Live-streaming platforms, particularly Twitch, have revolutionized online socialization and information dissemination. Twitch, initially designed for video game streaming, now hosts diverse content including political, leisure, and scientific discussions. This study explores climate change discourse on Twitch, a relatively under-researched area. Previous research on climate change on social media largely focuses on platforms like Twitter, Facebook, and Reddit, often concentrating on user reactions rather than video content itself. This study addresses this gap by employing an objective, quantitative methodology to analyze both video streams and accompanying chat logs. This methodology leverages automatic speech recognition (ASR) and natural language processing (NLP) to analyze a large dataset comprising nearly 130 hours of video and over 150 GB of data, thereby overcoming the limitations of manual analysis and ensuring reproducibility. The study aims to provide a comprehensive characterization of climate change discourse on Twitch and to contribute to understanding how public discussions on climate change adapt and utilize technological platforms.

Existing research on climate change discourse on social media primarily focuses on Twitter, Facebook, and Reddit. Studies on Twitter often highlight polarization, misinformation, and echo chambers. Research on Reddit examines the influence of climate-related events on collective action and deliberation. YouTube studies frequently concentrate on user comments rather than video content, with some exceptions analyzing video content to assess alignment with scientific consensus. The lack of research on Twitch is attributed to the platform's relative novelty, its perceived relevance primarily to gaming culture, and the challenges of analyzing extensive video data. While some Twitch research exists on gaming-related topics, few studies utilize the platform to address broader societal questions. This study aims to fill this gap by providing a comprehensive analysis of climate change discourse on Twitch, focusing on both video content and user interactions.

This quantitative study analyzed 78 Twitch video streams and their associated chat logs, collected using keywords "climate change" and "global warming." The sample included 56 unique streamers, with most broadcasting solo. Content varied, encompassing gaming, news, politics, science, art, and economics. Data processing used the Python Twitch API and Apify platform for title scraping. Audio from each stream was transcribed using Google Cloud Speech API, and chat logs were processed to extract messages, unique users, emotes, and URLs. A network analysis was performed using the NetworkX Python package, creating an adjacency matrix representing streamers and their shared followers. Louvain's algorithm identified community structures, while the assortativity coefficient measured interaction between communities. For content analysis, transcripts underwent preprocessing (tokenization, lemmatization, stemming, stop word removal) before applying Latent Dirichlet Allocation (LDA) to identify topics. The Stanford Named Entity Recognition (NER) tagger, combined with the Media Bias Fact Check (MBFC) database, assessed the ideological bias and credibility of cited media sources. A Context Index (CI) was created to evaluate the formality of each stream based on sentence length, formal language use, swearing, and stream categorization. Finally, sentiment analysis using VADER assessed viewer reactions and opinions, integrating NER to analyze sentiments toward specific named entities.

Network analysis revealed three major communities: Talk & Play (TP), Science & Education (SE), and Podcast & Politics (PP). TP included the most influential streamers, while SE and PP were similarly sized and interconnected, suggesting cross-community engagement. The assortativity coefficient (-0.28) indicated a non-polarized network. Analysis of chat logs showed a wide range of interaction levels, with a few streams achieving high impact, primarily political talk shows. Sentiment analysis revealed predominantly positive and neutral feelings, with negative comments generally low except in cases of controversial topics. Analysis of top named entities revealed mixed sentiments. Topic modeling identified eleven major topics, with climate change discussion and US politics being prominent, while climate skepticism was relatively low. Analysis of media sources showed a prevalence of trusted, left-leaning and centrist sources, with a minority of streams utilizing questionable sources. The Context Index revealed a relatively even distribution of formal and informal streams.

The findings challenge prior research suggesting polarization and echo chambers on other social media platforms. Twitch demonstrates a less polarized environment with significant cross-community engagement. While some highly influential streamers utilize questionable sources, the overall prevalence of misinformation appears relatively low. The high degree of connectivity between communities, as measured by the assortativity coefficient and the proximity matrix, suggests that the platform does not exhibit strong interactional polarization. The study also highlights the co-existence of formal and informal communication styles on Twitch, indicating its capacity to support both casual conversation and structured discussions. This is further emphasized by the relatively equal distribution between formal and informal streams. The platform facilitates both casual and formal discussions about climate change.

This study provides novel insights into climate change discourse on Twitch, highlighting its potential as a valuable site for understanding young people's attitudes. The findings challenge existing assumptions about online political polarization and misinformation. The methodology presented offers a valuable approach for analyzing large-scale video data. Future research should explore the longitudinal dynamics of climate change discourse on Twitch and investigate the impact of visual content on shaping discussions. Further research could also explore the evolution of the community and the development of new types of online discussions.

The study's limitations include the lack of video analysis, which could provide additional contextual information. The temporal validity of the findings is limited by the ephemeral nature of Twitch content. The relatively small sample size may restrict the generalizability of results. Future research should address these limitations by incorporating video analysis, conducting longitudinal studies, and expanding the sample size.