The final image cannot reveal where bias entered the pipeline
An international research team posted the WORLDVIEW study on September 10. Commercial image services may rewrite a user's request into a more detailed internal prompt, while users often cannot inspect or disable that step. The researchers created 8,960 prompts across 15 languages and 31 language-context pairings, then audited revisions produced in DALL‑E 3, Imagen 4 and GPT‑Image‑1.5.
The paper reports that US context was marked least relative to a context-free English baseline, while non-Western and non-Anglophone contexts received more explicit cultural additions. Diverse topics were flattened into narrow recurring vocabularies and recognizable stereotypes such as traditional symbols and settings. By sending original and revised prompts to models without their own revision layer, the team argues that rewriting itself can causally change the cultural representation in the image.
Three systems cannot represent every service and culture
This is a preprint awaiting peer review. Results can change with product versions, safety policies, translation choices and undisclosed pipeline updates. A few quantitative measures cannot fully capture cultural diversity, and revision can also improve safety or clarify an instruction.
The practical recommendation is auditability rather than automatically eliminating all rewriting. Services can explain the difference between user text and model input, give creators a way to review or correct cultural details, and test the complete deployed pipeline—not only the base model—across languages.