Harm or Humor (English Images) - Macro-F1: leaderboard
Metric: Macro-F1 (%) over the safe and harmful classes on the 3,701 English image jokes (memes); each joke is labelled safe or harmful (harmful jokes being explicit or implicit) and the model classifies it as safe or harmful zero-shot; blocked or empty responses count as harmful; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 72 | #105 |
| 2 | GPT-4o | 70.8 | #333 |
| 3 | Gemini 2.5 Pro | 67.9 | #145 |
| 4 | Qwen 2.5 VL 32B Instruct | 67.4 | #443 |
| 5 | InternVL2-8B | 55.8 | #826 |
| 6 | Gemini 3 Pro | 55.7 | #77 |
| 7 | Llama 3.2 11B Instruct | 42.8 | #1112 |
| 8 | aya-vision-8B | 39.5 | #1094 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=harm-or-humor-english-images-macro-f1 · How It Works · Data refreshed daily, snapshot 2026-10-11.