Harm or Humor: A Multimodal, Multilingual Benchmark for Overt and Covert Harmful Humor

2026-03-18

AI Summary

This work addresses the limitations of existing static benchmarks in detecting harmful humor that relies on cultural context and implicit cues, as well as their inability to distinguish between explicit and implicit harm. To bridge this gap, the authors introduce the first multimodal, multilingual benchmark supporting fine-grained classification into safe, explicitly harmful, and implicitly harmful categories across text, images, and videos. The benchmark spans English, Arabic, and culturally general contexts, with expert human annotations ensuring cultural sensitivity and support for deep reasoning. Systematic evaluation of leading large language models reveals that closed-source models significantly outperform open-source counterparts and exhibit notable performance disparities across languages, underscoring the critical role of cultural awareness and reasoning capabilities in effective safety alignment.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Dark humor often relies on subtle cultural nuances and implicit cues that require contextual reasoning to interpret, posing safety challenges that current static benchmarks fail to capture. To address this, we introduce a novel multimodal, multilingual benchmark for detecting and understanding harmful and offensive humor. Our manually curated dataset comprises 3,000 texts and 6,000 images in English and Arabic, alongside 1,200 videos that span English, Arabic, and language-independent (universal) contexts. Unlike standard toxicity datasets, we enforce a strict annotation guideline: distinguishing \emph{Safe} jokes from \emph{Harmful} ones, with the latter further classified into \emph{Explicit} (overt) and \emph{Implicit} (Covert) categories to probe deep reasoning. We systematically evaluate state-of-the-art (SOTA) open and closed-source models across all modalities. Our findings reveal that closed-source models significantly outperform open-source ones, with a notable difference in performance between the English and Arabic languages in both, underscoring the critical need for culturally grounded, reasoning-aware safety alignment. \textcolor{red}{Warning: this paper contains example data that may be offensive, harmful, or biased.}

Problem

• harmful humor
• multimodal
• multilingual
• implicit toxicity
• safety alignment

Innovation

• multimodal benchmark
• harmful humor detection
• implicit vs explicit toxicity
• multilingual safety evaluation
• contextual reasoning