Commercial text-to-image systems rewrite user prompts before generating images, and users cannot see or disable the rewrite. The authors introduce WORLDVIEW, a multilingual benchmark of 8,960 prompts spanning 15 languages and 31 language-context pairings, and use it to audit the revision layer in DALL-E-3, Imagen-4, and GPT-Image-1.5. They measure how strongly each cultural context is marked, whether it is flattened into a narrow vocabulary, and whether that vocabulary is stereotypical.
Compared to an English no-context baseline, US contexts are marked least, while non-Western and non-Anglophone contexts are marked much more heavily, flattened, and reduced to recognizable stereotypes. By comparing images from original and revised prompts on models without a revision layer, the authors show the revision layer itself is a causal source of stereotyping. This matters because bias audits that only look at final images attribute the bias to the image generation model, and fixes aimed at that model may miss the origin of the problem.