Toxicity in Twitch Chats: An LLM-Based Analysis Across Gaming Communities

Technical Summary: Toxicity in Twitch Chats: An LLM-Based Analysis Across Gaming Communities

Problem Statement
Online gaming communities, particularly on streaming platforms like Twitch, are frequently plagued by toxic behavior, including harassment, discrimination, and abusive language. While toxicity is a recognized challenge, existing research often focuses on in-game interactions or lacks granularity regarding how toxic behavior varies across different gaming communities and genres. Current moderation efforts are largely reactive, and the subjective, context-dependent nature of toxicity makes it difficult to define and detect consistently. There is a need for empirical insights into how toxicity manifests across specific games and genres to inform more targeted moderation strategies.

Methodology
The authors present a large-scale analysis of approximately 20.2 million chat messages collected from 4,452 Twitch streams. The study focuses on the two most popular games within seven selected Twitch categories, spanning a time period from June 2024 to November 2025.

• Data Collection and Preprocessing: Data was sourced via SullyGnome and downloaded using TwitchDownloaderCLI. The dataset includes 29,001 hours of stream time. To optimize computational resources, the authors pre-labeled objectively non-toxic messages (e.g., "hi," "gg," bot messages) and excluded them from the classification pipeline.
• Classification Approach: The study employs a zero-shot classification approach using a pre-trained Large Language Model (LLM), specifically the 14B-parameter Phi4 model. This model was chosen for its balance of performance and computational efficiency, as well as its lower tendency to refuse processing abusive content compared to models like Llama3.
• Taxonomy: Messages were classified according to Twitch's official moderation taxonomy, which comprises four main categories (Harassment, Discrimination, Sexual Content, Profanity) and eight subclasses (e.g., Bullying, Aggression, Misogyny, Swearing).
• Contextual Awareness: To address the context-dependency of toxicity, the model was provided with a temporal context window of the preceding ten seconds of chat messages during inference.
• Validation: The pipeline's reliability was evaluated through a human-model agreement study (Cohen's Kappa) and a comparison against a fine-tuned RoBERTa model on the TextDetox dataset.

Key Contributions

1. Reproducible Pipeline: The paper provides a reproducible pipeline for toxicity detection and classification using zero-shot LLMs, achieving an F1 score of 94.5% on the TextDetox benchmark, slightly outperforming a fine-tuned RoBERTa baseline.
2. Human-Model Agreement: The study demonstrates that the LLM's agreement with human raters (Kappa ~0.53) is comparable to inter-human agreement (Kappa ~0.55), validating the model's utility for large-scale analysis despite the subjective nature of toxicity.
3. Granular Empirical Analysis: The work offers a detailed breakdown of toxicity patterns across 14 specific games and four distinct genres, moving beyond broad generalizations to identify game-specific and genre-specific toxicity profiles.

Results

• Overall Prevalence: Approximately 2.4% of all analyzed messages were classified as toxic.
• Dominant Toxicity Types: Harassment was the most prevalent primary category (75.6%), with bullying being the most common subclass. Profanity frequently appeared as a secondary label (53.0%), often intensifying other toxic behaviors.
• Game-Level Variations: Toxicity rates varied significantly by game. Red Dead Redemption II exhibited the highest toxicity rate (3.96%), while Minecraft showed the lowest (0.74%). Notably, Counter-Strike 2 and Valorant displayed similar toxicity distributions, likely due to shared competitive structures, whereas other games within the same broad categories showed distinct profiles.
• Genre-Level Variations: Multiplayer Online Battle Arena (MOBA) streams had the highest relative toxicity rate (3.2%), followed by Single-Player Shooters (3.1%). Sports games had the lowest (2.0%).
• Distribution Differences: While harassment and profanity were common across all genres, Single-Player Shooter streams showed higher proportions of aggression, misogyny, and sexual content compared to other genres. Statistical analysis (PERMANOVA) indicated significant differences in toxicity distributions between most games and genres, though dispersion analysis (PERMDISP) suggested that variability within genres also plays a significant role.

Significance and Claims
The authors claim that their findings provide empirical evidence that toxicity is not merely a platform-wide phenomenon but is heavily shaped by game-specific community norms and mechanics. The study suggests that:

• Context Matters: Toxicity patterns differ significantly even between games within the same genre, implying that "one-size-fits-all" moderation strategies may be less effective than those tailored to specific game communities.
• Moderation Gaps: The presence of toxic content in the dataset (messages that survived platform filtering) highlights limitations in current automated moderation systems.
• Taxonomy Limitations: The study notes that the ambiguity between certain categories (e.g., harassment vs. profanity) in Twitch's taxonomy contributes to classification disagreements, suggesting that clearer definitions could improve consistency.

The paper concludes that while genre provides a useful analytical lens for broad tendencies, the specific game and its surrounding community are the primary drivers of fine-grained toxicity variations. These insights aim to support the development of more targeted and effective moderation strategies for gaming communities.