HatefulStoryPrompts: New Benchmark Exposes Hateful Narratives in Multi-Turn Visual Story Generation

A recent paper published on arXiv (2608.05210) presents HatefulStoryPrompts, a benchmark dataset aimed at assessing hateful intent in multi-turn visual story generation by cutting-edge text-to-image (T2I) systems. The research notes that picture books and comics have long been vehicles for disseminating hateful messages, referencing the infamous Nazi propaganda book 'Der Giftpilz.' With the rise of sophisticated T2I systems like Gemini and GPT-Image, which can maintain character and scene consistency throughout dialogues, the production of hateful visual stories—sequences of images that together express hateful themes—has become both cost-effective and scalable. While earlier studies concentrated on individual images, the group-level hateful meaning had not been thoroughly examined. To fill this void, the researchers developed HatefulStoryPrompts, consisting of 330 multi-turn setups based on 55 hateful stories in two languages and three visual styles. They tested five leading models across 4,950 attempts, confirming that all models successfully completed the tasks, although the abstract does not specify the results. The research emphasizes the risk of generative AI being misused to create harmful visual narratives and advocates for further exploration into detecting and mitigating group-level hateful content.