MBZUAI's Harm or Humor Benchmark Exposes Frontier AI's Cultural Blind Spot
A new benchmark tests whether frontier models can spot dark jokes across English, Arabic, images, and video, and the results expose a cultural reasoning gap.
ProTakeaways
- New Harm or Humor benchmark covers 3,000 texts, 6,000 images, and 1,200 videos in English and Arabic.
- Labels split jokes into safe, explicit harmful, and implicit harmful to test cultural reasoning.
- GPT-5.2 leads text at 90.2 F1 English and 83.1 F1 Arabic, still drops 14 points on implicit Arabic harm.
- Open-source vision models like LLaVA-NeXT and Aya collapse to near-zero harmful detection due to over-alignment.
- GPT-5 Pro tops video at 80.2 F1, but Gemini-3-Pro underperforms its predecessor Gemini-2.5-Pro.
- Dataset released CC BY-NC-SA 4.0 for research; no model weights, taxonomy is the contribution.
A joke can be a joke, or it can be a slur wrapped in a punchline. Telling the two apart is a safety problem most toxicity benchmarks skip, because they lean on explicit slurs and profanity rather than the subtle, culturally loaded humor that actually circulates online. A new benchmark called Harm or Humor tries to close that gap, and its results suggest even frontier models are still leaning on surface cues.
Built by researchers at MBZUAI and INSAIT, the benchmark evaluates language, vision-language, and video models on their ability to distinguish safe jokes from harmful ones, then further splits the harmful cases into overt versus covert offense. The manually curated dataset comprises 3,000 texts and 6,000 images in English and Arabic, alongside 1,200 videos spanning English, Arabic, and language-independent (universal) contexts.
Why existing safety tests miss the joke
The authors argue that current benchmarks have three blind spots. Most focus on text or single images and ignore video. They skew heavily English, with little coverage of Arabic or language-independent visual humor. And they rarely isolate implicit harm, where the offense only surfaces after cultural or semantic reasoning.
To force models past pattern matching, the team defined a strict labeling scheme:
- Safe: no sensitive themes involving violence, race, disability, religion, or sexuality.