Art, Censorship, and AI: Investigating the Role of Large Language Models in Artistic Freedom

This dataset accompanies a study on how large language models interpret and evaluate censorship in visual art across different national contexts. The project investigates whether AI systems reflect, reproduce, or diverge from real-world censorship norms when assessing politically, religiously, and socially sensitive artworks. The repository contains: Image Stimuli (n=100)A curated set of artworks covering themes such as political critique, religion, nudity, sexuality, violence, and death. The images are organised into thematic folders. Model OutputsStructured responses in CSV format from four multimodal large language models: OpenAI (GPT-5) Mistral Pixtral (Mistral Small 4) Qwen (qwen-vl-max) HyperCLOVA X (HyperCLOVAX-SEED-Vision-Instruct-3B) Each model was prompted to evaluate whether an artwork would likely be censored in national museum contexts across multiple countries, and to provide justification and confidence scores. Prompt SpecificationThe exact system and user prompts used across all models are included to ensure full reproducibility. Contributor Roles:AC: Conceptualisation; Data Curation; Investigation; Methodology; Validation.AF: Conceptualisation; Methodology; Software.