When Jokes Cross the Line: Analyzing Regular Humor and Dark Humor in YouTube Shorts

Researchers introduce TwistedHumor, a dataset of 1,211 YouTube Shorts with 33,041 annotated comments, to study the boundary between acceptable humor and harmful content on short-form video platforms. The analysis reveals that dark humor clusters around critique and coping themes, generates more mixed audience reactions than regular humor, and exposes limitations in current large language models for content moderation tasks.

The TwistedHumor dataset addresses a critical gap in content moderation research by providing empirical evidence of how humor operates in the gray zone between entertainment and potential harm on social platforms. YouTube Shorts and similar short-form video formats have become dominant content distribution channels, yet academic understanding of how audiences interpret contextual nuance in brief, highly engaging clips remains limited. This research matters because platform moderation teams increasingly rely on automated systems that struggle with context-dependent judgments about acceptable speech.

The study's multi-view analysis reveals that dark humor serves distinct communicative functions—enabling social critique, facilitating coping mechanisms, and exploring identity—rather than constituting a monolithic category. Comment sentiment analysis shows regular humor generates predominantly positive responses while dark humor receives more neutral and toxic reactions, suggesting audience interpretation varies significantly based on humor type. These findings indicate that blanket content policies fail to account for the nuanced ways users deploy different humor styles.

For platform developers and policy makers, this research demonstrates that effective moderation requires context-aware approaches distinguishing between harmful content and humor that pushes boundaries intentionally. The evaluation of large language models against human annotations reveals performance gaps, particularly with brief jokes versus stand-up comedy, highlighting current AI limitations in nuanced content assessment. This benchmark enables developers to test and improve moderation systems before deployment.

Looking forward, platforms must invest in training data that captures humor's contextual dimensions and audience reception patterns. The dataset methodology provides a template for studying similar gray-area content across different platforms and demographics, potentially informing more sophisticated, culturally-aware moderation standards.

• →Dark humor clusters around critique, coping, and identity expression rather than forming a uniform harmful category
• →Regular humor receives more positive audience sentiment while dark humor generates more mixed and occasionally toxic reactions
• →Current large language models perform significantly better on stand-up comedy than on brief social media jokes
• →The TwistedHumor dataset enables researchers to study the boundary between acceptable and harmful content on short-form platforms
• →Context-aware moderation approaches are essential, as blanket policies fail to account for intentional boundary-pushing humor