💬 NLP

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

This paper introduces a novel multimodal and multilingual benchmark comprising 3,000 texts, 6,000 images, and 1,200 videos in English and Arabic to evaluate the ability of state-of-the-art models to distinguish between safe, overtly harmful, and covertly harmful humor, revealing significant performance gaps between closed- and open-source models and highlighting the urgent need for culturally grounded safety alignment.

Original authors: Ahmed Sharshar, Hosam Elgendy, Saad El Dine Ahmed, Yasser Rohaim, Yuxia Wang

Published 2026-03-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ahmed Sharshar, Hosam Elgendy, Saad El Dine Ahmed, Yasser Rohaim, Yuxia Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to understand human jokes. You show it a picture of a cat wearing a hat, and it laughs. Easy! But then you show it a joke that sounds funny on the surface but is actually mean-spirited, or a meme that relies on a specific cultural inside joke. Suddenly, the robot gets confused. It might think a harmless joke is dangerous, or worse, it might laugh at a joke that is actually hurtful.

This paper, "Harm or Humor," is like a giant, tricky "final exam" designed to test exactly how good AI is at spotting the difference between a funny joke and a harmful one.

Here is the breakdown of the paper using some everyday analogies:

1. The Problem: The Robot's "Cultural Blindness"

Think of AI models as newly hired interns who have read every book in the library but have never actually lived in the real world. They know the dictionary definitions of words, but they don't understand nuance.

The "Overt" Harm: This is like a clown wearing a sign that says "I am being mean." The AI can easily spot this because the "mean" part is written in big, bold letters.
The "Covert" (Implicit) Harm: This is like a clown whispering a joke that only people from a specific village understand. To an outsider, it sounds innocent. To the villagers, it's a cruel insult. Current AI models often miss these whispers because they lack cultural context and the ability to "read between the lines."

2. The Solution: The "Harm or Humor" Exam

The researchers created a massive, three-part test to see if AI can pass the "Cultural Intelligence" course. They didn't just use text; they used a mix of media, just like how we consume humor in real life.

The Text Section: 3,000 jokes (in English and Arabic).
The Image Section: 6,000 memes (pictures with text).
The Video Section: 1,200 short clips (where timing, voice, and visuals all matter).

The Twist: They didn't just ask, "Is this funny?" They asked, "Is this safe?" And if it's not safe, is the danger obvious (like a slapstick fall) or hidden (like a subtle stereotype)?

3. The Test Subjects: The "Students"

They tested two types of AI "students":

The Private Schools (Closed-Source Models): These are the big, expensive AI models from companies like OpenAI (GPT) and Google (Gemini). They have seen a lot of data and have strict rules.
The Public Schools (Open-Source Models): These are models anyone can download and study. They are often more flexible but sometimes less disciplined.

4. The Results: Who Passed?

The results were a bit of a shock, like a student acing a math test but failing a test on local customs.

The "English" Bias: The AI models were much better at understanding jokes in English than in Arabic. It's like a student who is fluent in French but stumbles over simple Spanish idioms. Even the smartest models struggled with the cultural nuances of Arabic humor.
The "Implicit" Trap: The models were great at spotting the "clowns with signs" (Explicit harm). But when it came to the "whispered jokes" (Implicit harm), they often failed. They missed the hidden insults because they were looking for surface-level keywords rather than deep meaning.
The "Safety Overreaction": Some open-source models were so scared of making a mistake that they refused to answer or labeled everything as "safe." It's like a security guard who, afraid of letting a bad guy in, locks the door and keeps everyone out. This is bad because it means they miss the actual harmful jokes.

5. The Big Takeaway

The paper concludes that making AI bigger isn't enough. You can't just feed an AI more data and expect it to understand human culture.

To make AI safe, we need to teach it cultural empathy. We need it to understand why a joke hurts, not just what words are used. Currently, AI is like a tourist who knows the map but doesn't know the local customs; it needs to learn how to be a local to truly understand humor and safety.

In short: The paper built a tough test to show that while AI is getting smarter, it still struggles to understand the "unspoken rules" of humor, especially in languages and cultures other than English. Until we fix this, AI might keep accidentally laughing at things it shouldn't, or getting offended by things it shouldn't.

1. Problem Statement

Current AI safety benchmarks fail to adequately address the complexities of dark humor, particularly the distinction between explicit (overt) toxicity and implicit (covert) harm.

The Challenge: Dark humor often relies on subtle cultural nuances, world knowledge, and multi-step reasoning to decode harmful intent. Unlike standard toxicity detection (which looks for slurs or graphic violence), implicit humor requires understanding context, cultural norms, and semantic inference.
The Gaps: Existing datasets suffer from three main limitations:
1. Modality Bias: Most focus on static text or single images, neglecting the temporal and multimodal nature of video humor.
2. Language Gap: There is a heavy bias toward English, with a lack of resources for low-resource languages like Arabic and language-independent (universal) visual humor.
3. Implicit Harm Neglect: Few benchmarks explicitly categorize and evaluate the detection of implicit harmful content, which is the most difficult safety challenge for current models.

2. Methodology

A. Dataset Construction ("Harm or Humor")

The authors curated a novel, manually annotated dataset comprising 10,207 items across three modalities:

Text: 3,000 jokes (2,000 English, 1,000 Arabic).
Images: 6,005 memes (3,701 English, 2,304 Arabic).
Video: 1,202 clips (English, Arabic, and Universal/language-independent).

Annotation Taxonomy:
The dataset uses a strict, hierarchical labeling scheme:

Safe vs. Harmful: Determined by whether the content uses sensitive themes (violence, racism, sexuality, disability, religion) to demean or target groups.
Explicit vs. Implicit (for Harmful only):
- Explicit: Overt toxicity (e.g., profanity, graphic imagery) identifiable without deep reasoning.
- Implicit: Covert toxicity requiring semantic understanding, cultural context, and multi-step reasoning to decode.

Curation Process:

Sources: Manually collected from online forums, social media (Reddit, X), and existing repositories (e.g., D-HUMOR). No automated scraping was used to ensure quality.
Annotators: Seven volunteer annotators (diverse backgrounds, native Arabic speakers with various dialects, and fluent English speakers) labeled the entire dataset.
Consensus: Final labels were determined via majority voting. Inter-annotator agreement was high (Fleiss' $\kappa$ > 0.80 for text, >0.74 for video).

B. Evaluation Framework

The benchmark evaluates State-of-the-Art (SOTA) models across all modalities using a unified binary classification task (Harmful vs. Safe).

Models Evaluated:
- Closed-Source: GPT-4o, GPT-5 series (5.2, 5 Pro), Gemini 2.5 Pro, Gemini 3 Pro.
- Open-Source: Llama-3.1, DeepSeek-Reasoner, Qwen2.5 series, AceGPT, ALLaM, Jais, and various Vision-Language Models (VLMs) like LLaVA-NeXT and InternVL.
Metrics:
- Accuracy & Macro-F1: Overall performance.
- Recall (True Positive Rate): Specifically for Implicit and Explicit subsets. The authors emphasize recall because missing harmful instances (false negatives) is the critical failure mode in safety alignment.
Prompting: Unified prompts were used across languages and modalities to ensure fair comparison.

3. Key Contributions

First Multimodal, Multilingual Benchmark for Implicit Harm: Introduces a dataset spanning text, images, and videos in English, Arabic, and Universal contexts, specifically targeting the gap between overt and covert harmful humor.
Cultural Grounding: Addresses the English bias in existing humor datasets by including diverse Arabic dialects and language-independent visual content, highlighting the need for culturally aware safety alignment.
Systematic Cross-Modal Evaluation: Provides a comprehensive analysis of how open-source vs. closed-source models perform across different modalities and languages, revealing specific failure modes in reasoning and cultural understanding.

4. Key Results and Analysis

A. Closed-Source vs. Open-Source Performance

Closed-Source Dominance: Models like GPT-5.2 and Gemini-2.5 Pro significantly outperform open-source models.
- Text: GPT-5.2 achieved 90.2% F1 (English) and 83.1% F1 (Arabic).
- Image: GPT-5.2 led with 72.0% F1 (English).
- Video: GPT-5 Pro led with 80.2% F1.
Open-Source Struggles: Open-source models generally lag, particularly in Arabic and implicit detection.
- Safe Bias: Many open-source VLMs (e.g., LLaVA-NeXT, Aya-Vision) exhibit a severe "safe bias," defaulting to "Safe" or refusing to answer, resulting in near-zero recall for harmful content.
- Reasoning Gap: Open-source models struggle significantly with implicit harm, often failing to decode the underlying cultural or semantic cues.

B. The Language Gap (English vs. Arabic)

Performance Drop: All models show a systematic performance drop when moving from English to Arabic.
- Example: GPT-5.2's recall for implicit harm in Arabic dropped from 71.9% (Explicit) to a lower figure, showing a ~15-22% drop in performance compared to English.
Cultural Nuance: Models trained primarily on English data fail to recognize culturally specific implicit harms in Arabic dialects.
Specialized Models: Arabic-specific models (e.g., ALLaM, Jais) showed high accuracy in detecting implicit Arabic humor (up to 98% recall) but suffered from high false-positive rates (low overall accuracy), indicating they are overly sensitive.

C. Modality-Specific Findings

Text: Models perform best here but still struggle with implicit puns and cultural references.
Images: The gap between English and Arabic is widest here. Models fail at OCR and dialectal text extraction in Arabic memes, leading to a collapse in detection capabilities.
Video: This is the most challenging modality. While GPT-5 Pro leads, open-source models (like VideoChat) often fail completely on Arabic content (0.0% F1), effectively acting as monolingual systems despite multimodal architectures.

D. Implicit vs. Explicit Detection

Reliance on Surface Cues: Most models rely on explicit markers (e.g., visible weapons, slurs). When these are absent (implicit), performance drops sharply.
Reasoning Deficit: The inability to perform multi-step reasoning to decode the "punchline" of a harmful joke is the primary bottleneck.

5. Significance and Implications

Safety Alignment Limitations: The results demonstrate that simply scaling model size is insufficient for safety. Current models lack culturally grounded reasoning capabilities required to detect implicit harm.
Need for Cultural Calibration: Safety guardrails must be language-specific and culturally aware. A global threshold for "harm" fails in multilingual contexts, leading to the under-moderation of harmful content in low-resource languages like Arabic.
Future Directions: The paper calls for:
- Reasoning-aware alignment techniques that force models to articulate why content is harmful.
- Development of lightweight, open-source architectures capable of integrating audio, visual, and temporal signals for video analysis.
- Expansion of benchmarks to include finer-grained harm-type taxonomies (e.g., specific types of racism or sexism).

Conclusion: The "Harm or Humor" benchmark reveals a critical gap in AI safety: while models can detect obvious toxicity, they are largely blind to the subtle, culturally dependent, and implicit nature of harmful humor, especially in non-English languages. This underscores the urgent need for more sophisticated, reasoning-based safety alignment strategies.