Arabic memes and prompts harmful content detection results revealed
ArGuard Shared Task: Harmful Content Detection in Arabic Memes and LLM Prompts
Computation and LanguageArtificial Intelligence
Summary
Detecting harmful or hateful content in Arabic memes and text prompts used in large language models is challenging. The authors organized a competition with two tasks: one to spot hateful images with text, and another to find unsafe text prompts in Arabic. Many teams participated using different AI models, with varied success. The hardest challenge was identifying detailed types of hate in memes due to little training data and differences between training and testing examples.
What this means in practice
- •For social media moderators: Use tested models and benchmarks to improve detection of hateful Arabic memes combining images and text for safer online communities.
- •For ai safety engineers: Evaluate and refine Arabic language model prompts to prevent generation of harmful content using evaluated harmful prompt detection approaches.
Authors
Firoj Alam, Md. Rafiul Biswas, Mohamed Bayan Kmainasi, Ali Ezzat Shahroor, Hamdy Mubarak, George Mikros, Abul Hasnat, Wajdi Zaghouani
Abstract
ArGuard is a shared task on harmful content detection in Arabic memes and LLM prompts. It includes two tracks: Track A focuses on multimodal hate detection in Arabic memes, while Track B addresses harmful prompt detection for Arabic LLM safety evaluation. In total, 58 teams registered, 35 participated in the final evaluation, and 27 submitted system-description papers. Participating teams explored models such as AraBERT, Jais, and Qwen3-VL. The best systems achieved macro-F1 scores of 0.823 on A1, 0.419 on A2, 0.984 on B1, and 0.790 on B2. Fine-grained meme classification in A2 was the most challenging setting, partly due to sparse labels and train-test distribution shifts.