What they did

The authors audited the hidden prompt-revision layers of DALL-E-3, Imagen-4, and GPT-Image-1.5 using WORLDVIEW. The benchmark contains 8,960 prompts across 15 languages and 31 language–context pairings, with no-context English prompts serving as a baseline.

They measured how strongly revisions added cultural markers, whether diverse requests were rewritten using a narrow recurring vocabulary, and whether those additions reflected stereotypes. To isolate causality, they also generated images from original and revised prompts using models that did not have their own revision layer.

Key findings

• The United States was the least-marked cultural context relative to the no-context English baseline; non-Western and non-Anglophone contexts received substantially more added cultural description.
• Prompt revisions compressed culturally situated requests into narrow vocabularies that recurred across otherwise diverse topics.
• The recurring descriptors often invoked recognizable cultural stereotypes rather than preserving the breadth of the original requests.
• Comparing images produced from original versus revised prompts showed that revision itself can cause stereotypical visual outputs, rather than merely reflecting biases arising later in image generation.

Why it matters

The study shows that cultural bias can enter a text-to-image product through its surrounding software pipeline, not only through the image model. Audits and mitigations therefore need access to deployed prompt transformations; evaluating final images or base models alone may misidentify where biased representations originate.

Caveats

The audit covers three commercial systems and a fixed benchmark, so its conclusions may not generalize to every provider, language, cultural setting, or future system version. The abstract also does not report effect sizes or the reliability of stereotype judgments, and proprietary revision mechanisms remain opaque beyond their observable outputs.