PaperlayerView exact revision on arXiv

Source-linked AI summary

Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems

Aleksandra Urman, Elsa Lichtenegger, Salima Jaoua, Azza Bouleimen, Robin Forsberg, Corinna Hertweck, Stefania Ionescu, Nicolò Pagan, Ancsa Hannak, Joachim Baumann

arXiv:2609.11532v1cs.AI

TL;DR

Existing audits of cultural bias in text-to-image systems examine final images without isolating the commercial prompt-revision layer. This paper introduces WORLDVIEW to audit that layer across languages and contexts, finding that it asymmetrically marks and stereotypes non-Western contexts and is causally linked to stereotypical visual outputs.

  • Problem

    Prior audits treat text-to-image generation as a unified pipeline, so they cannot determine whether cultural bias originates in prompt revision or visual generation.

  • Method

    The paper introduces WORLDVIEW and audits three commercial systems using markedness, cultural-flattening, stereotype, and controlled-ablation analyses.

  • Results

    The US is closest to the unmarked default, while non-Western contexts receive heavier marking, narrow stereotypical vocabularies, and causally linked stereotypical visual outputs.

  • Takeaways & Limitations

    Bias audits and mitigation should examine deployed systems, including prompt revision, rather than evaluating the image model alone.

  • Takeaways & Limitations

    WORLDVIEW underrepresents many regions, so its specific CMS and CFS rankings should not be generalized to unexamined contexts.

Abstract

from arXiv · show

Commercial text-to-image systems silently revise user prompts before generating images, a step users typically cannot disable or even see. Yet, existing audits of cultural bias examine only the final images and treat generation as a single pipeline, so they cannot tell where the bias originates. We introduce WORLDVIEW, a multilingual benchmark of 8,960 prompts across 15 languages and 31 language-context pairings. Using it, we audit the revision layer in three systems (DALL-E-3, Imagen-4, GPT-Image-1.5) through a three-step analysis of how heavily it marks each cultural context, whether it flattens that context into a narrow vocabulary, and whether that vocabulary is stereotypical. Relative to a no-context English baseline, the US is the least-marked context, while non-Western and non-Anglophone contexts are marked far more heavily, flattened into narrow vocabularies applied across topically diverse prompts, and reduced to recognizable cultural stereotypes. Comparing images from original versus revised prompts on models without a revision layer, we identify the layer itself as a previously undocumented, causal source of this stereotyping. To locate cultural bias, and fix it, we must audit the system as deployed, not the model alone.

1 Introduction

Existing audits usually treat text-to-image generation as a single pipeline, leaving prompt revision as an unseen potential source of cultural bias. This work introduces WORLDVIEW and audits that intermediate layer across three commercial systems.

  • 1 Introduction: Commercial systems revise user prompts before image generation, often without user awareness or the ability to disable the process.Revisions may expand, translate, or otherwise modify the original prompt.
  • 1 Introduction: Audits of final images alone cannot determine whether stereotypical representations originate in prompt revision or visual generation.This uncertainty can make image-model-only mitigation ineffective when bias is already encoded in revised prompts.
  • 1 Introduction: WORLDVIEW contains 8,960 prompts across 15 languages and 31 language–context pairings, with revised prompts, metrics, and evaluation tools released for auditing.The benchmark is designed to support audits of cultural and other bias in both T2I systems and their revision layers.
  • 1 Introduction: Auditing DALL-E-3, Imagen-4, and GPT-Image-1.5 shows that prompt revision is not neutral: US-centric representations remain closest to the unmarked default, while other contexts receive stereotypical, culturally flattened representations.The study also compares original and revised prompts in a controlled ablation to connect revision-layer changes with final visual outputs.

2 Related Work

Prior work documents cultural and demographic bias in text-to-image systems but generally audits only final visual outputs. This paper frames prompt revision as a separable source of bias and distinguishes cultural flattening from stereotyping.

  • 2 Related Work: Prior studies cover demographic stereotyping, Western-centric defaults, cultural misrepresentation, and multilingual disparities in T2I outputs.These studies commonly evaluate domains such as food, landmarks, and region-specific representation.
  • 2 Related Work: Treating the T2I pipeline as monolithic prevents attribution of stereotypical images to training data, prompt revision, image generation, or their interaction.This attribution problem limits the effectiveness of mitigation efforts.
  • 2.2 Prompt markedness: Markedness describes asymmetric treatment in which one category functions as an unmarked default while others receive explicit linguistic emphasis.The paper applies this framework to compare context-specified revisions with a no-country English baseline.
  • 2.3 Cultural flattening and stereotyping: Cultural flattening is a narrow representation of culture, whereas stereotyping is flattening that uses a specific set of stereotypical descriptors.The paper measures flattening with CFS and then qualitatively examines whether the resulting vocabulary is stereotypical.

3 Constructing WORLDVIEW

WORLDVIEW tests cultural adaptation using everyday prompts with unmarked English baselines and context-specified variants across multilingual language–context pairings. Its design spans diverse social and visual domains and uses native-speaker verification.

  • 3 Constructing WORLDVIEW: WORLDVIEW starts with 280 English prompts covering everyday situations and creates baseline and context-specified variants for each prompt.Baseline prompts omit geographic specification, while context variants add an explicit geographic reference in the associated language.
  • 3 Constructing WORLDVIEW: The baseline prompts span 14 domains, including family, transportation, public services, holidays, housing, politics, religion, policing, beauty, advertising, and immigration.The domains were selected because their visual representation can vary substantially across contexts.
  • 3 Constructing WORLDVIEW: The 280 prompts were translated into 14 additional languages, verified or corrected by native speakers, and paired with 31 national or regional contexts across 15 languages.For example, Spanish is paired with both Mexico and Spain.

4 Experimental setup

The study collects revised prompts and images from three commercial systems and audits revision through markedness, flattening, and stereotype analyses. It then links textual revisions to visual outputs using image descriptions, correlations, and a controlled ablation.

  • 4 Experimental setup: The audit submits 8,960 prompts to DALL-E-3, Imagen, and GPT-Image, yielding 8,808, 8,960, and 8,855 image–revised prompt pairs respectively.Differences reflect variation in safety-guardrail behavior after up to five retries.
  • 4.2.1 Step 1: Contextual Markedness Score: CMS measures semantic distance between context-specified and no-context English revisions, treating the latter as the unmarked baseline.It is descriptive rather than normative: high markedness indicates asymmetry in revision, not inherently problematic cultural differentiation.
  • 4.2.2 Step 2: Cultural Flattening Score: CFS measures cultural flattening by combining the prevalence of context-distinctive inserted terms with their spread across topically diverse prompts.Flattening requires both a distinctive vocabulary and broad deployment across prompts; the default implementation uses k = 10 terms and an arithmetic mean.
  • 4.2.3 Step 3: Stereotype analysis: The stereotype step qualitatively examines each context–model pair’s top 20 TF-IDF terms for recurring cultural tropes.This tests whether narrow vocabularies identified by CFS are also stereotypical.
  • 4.3 Isolating the causal link: The visual analysis uses Qwen2.5-VL descriptions as a lexical bridge, then combines cross-model correlations with an ablation comparing original and revised prompts in models without built-in revision.The correlation alone cannot establish causal direction, motivating the controlled ablation.

5 Results

The revision layer marks cultural contexts asymmetrically, flattens some into narrow vocabularies, and introduces stereotypical content that propagates into images. Comparisons of original and revised prompts identify revision as a causal source of these visual stereotypes.

  • 5.1 The US is closest to the unmarked default: CMS ranged from 0.21–0.31 for the US to as high as 0.47 for Finland+Finnish, with markedness generally increasing away from the Anglophone Western European default.The US was least marked across all three models, while Nordic and MENA contexts were at the high-markedness end.
  • 5.2 Swiss and Nordic contexts are flattened most: CMS and CFS captured different phenomena: Switzerland had moderate markedness but the highest flattening, whereas some Arabic-speaking contexts had high markedness but moderate flattening.Most contexts occupied either low-prevalence/high-spread or moderate-prevalence/low-spread regions; Imagen had the largest share of pervasive-flattening cases.
  • 5.3 Stereotypical content: High-CFS contexts were reduced to recognizable stereotypes, including Finland as winter imagery, Switzerland as Alpine tourism, Saudi Arabia as desert and religious markers, and Egypt as pharaonic imagery.These patterns recurred across all three models and across diverse prompt domains.
  • 5.4 Revision Layer is Causally Linked to Stereotyped Visual Outputs: Text-level CMS correlated positively with image-level CMS, with 91 of 93 context–model pairs significant and 33–46% of visually distinctive terms already distinctive in revised prompts.The correlations were ρ = 0.27 for DALL-E-3, 0.28 for GPT-Image, and 0.50 for Imagen.
  • 5.4 Revision Layer is Causally Linked to Stereotyped Visual Outputs: Revised-prompt images had higher CMS than original-prompt images, while mean CFS increased from 0.75 to 0.86 for SDXL and from 0.85 to 0.98 for Flux.Terms distinctive only under revised prompts were recognizably stereotypical, and 13–15 terms per model were significantly more prevalent after revision.

6 Discussion and Conclusion

The paper identifies prompt revision as an additional, previously unexamined source of cultural bias in deployed text-to-image systems. Its findings motivate auditing the full multi-stage pipeline and suggest that revision-stage interventions may address part of the bias.

  • 6 Discussion and Conclusion: The framework’s transferability beyond cultural bias is presented as a discussion implication rather than an evaluated result in the supplied passages.The supplied limitation passage only identifies transferability as a discussion topic.
  • 6 Discussion and Conclusion: The prompt revision layer asymmetrically marks non-Western contexts, compresses them into narrow stereotypical vocabularies, and is causally linked to stereotypical visual outputs.The discussion presents revision as an additional source beyond training data, architectures, and evaluation paradigms.
  • 6 Discussion and Conclusion: Audits should evaluate text-to-image systems as deployed rather than only individual generative models, because final-output audits conflate revision and image-generation stages.Separating these stages makes it possible to measure the revision layer’s independent contribution to cultural bias.
  • 6 Discussion and Conclusion: The framework can be transferred beyond cultural bias to social attributes such as gender by applying CMS, CFS, and the original-versus-revised ablation design.The supplied discussion states that the same metrics can test whether groups are marked more heavily or flattened.
  • 6 Discussion and Conclusion: Debiasing the image model alone may be limited when stereotypical content is introduced earlier, while revision-stage intervention could address part of the bias.The discussion qualifies that feasibility and sufficiency of intervention at the revision step remain open questions.

Limitations

The study’s limitations concern measurement, benchmark coverage, provider diversity, causal-analysis scope, pretrained image-model biases, and the choice to remain descriptive rather than normative.

  • VQA descriptions may introduce bias into image-content comparisons, although matched outputs and an informal inspection of 150 image–description pairs provided some reassurance.The authors did not observe VQA adding culturally relevant or stereotypical content absent from the images, but the check was not systematic.
  • The 31 language–context pairings underrepresent global cultural diversity, so specific CMS and CFS rankings should not be generalized to unexamined contexts.Sub-Saharan Africa, Southeast Asia, Central Asia, and much of Latin America are absent or sparsely represented.
  • All three audited systems come from OpenAI and Google, leaving the claim that patterns are not provider-specific in need of confirmation from additional providers.xAI could not be included because access to its revised prompts was deprecated.
  • The causal ablation covers four English-speaking contexts and GPT-Image, with only a limited Switzerland extension outside Anglophone contexts.The extension covers one context in three source languages and does not control for non-English-original images failing to represent prompts.
  • Open image models have their own biases, but matched original-versus-revised rendering holds those biases constant while estimating revision’s incremental effect.The authors nevertheless treat consistent increases in markedness and flattening under revised prompts as evidence attributable to the revision layer.
  • The analysis is descriptive rather than normative and does not specify what culturally appropriate, non-flattened representation should contain.The authors call for community-centered engagement rather than top-down specification by researchers.

Ethical Considerations

The paper addresses misuse, cultural authority, contested contexts, data handling, and the broader implications of targeting prompt revision as an intervention point.

  • The benchmark and toolkit could enable reverse-engineering or deliberate elicitation of stereotypical imagery, although the authors judge transparency benefits to outweigh this risk.The documented vocabularies are already being injected into user-facing image generations without public scrutiny.
  • The authors do not claim authority to speak for any of the 31 cultural contexts, despite diverse authorship and collaborative prompt design.They note that identifying stereotypes requires cultural knowledge and that their qualitative terms align with widely documented stereotypes.
  • Palestine and Taiwan are included for representational coverage, not as a political position on sovereignty, motivating the paper’s use of “context” rather than “country.”The same terminology accommodates contexts with contested political status.
  • The study collected data through official APIs, withheld generated images to reduce redistribution risks, and released revised prompts, metrics, and code for reproducibility.This release strategy aims to support auditing while minimizing harm.
  • Removing stereotypical vocabulary from revised prompts is necessary but insufficient because training data, model architecture, and evaluation paradigms also shape equitable representation.The paper therefore frames prompt revision as an actionable intervention point rather than a complete solution.
  • WORLDVIEW uses 15 languages and 31 language–context pairings selected for geographic, linguistic, and economic diversity within native-speaker and budget constraints.Some pairings reflect realistic multilingual use or colonial and post-imperial linguistic legacies.

B Detailed data collection procedures

The study collected revised prompts and images from three commercial systems while documenting model-specific access, generation, refusal, and budget conditions.

  • The three systems were chosen across providers, architectures, and collection periods, but inclusion depended on providers returning revised prompts.xAI was excluded after its revised-prompt access was deprecated.
  • DALL-E-3: DALL-E-3 translated non-English inputs to English and returned revised prompts through its API, with up to five images and repeated refusal attempts per combination.Images were generated at 1024×1024 resolution in standard quality.
  • Guardrails: Refusals were systematically concentrated in politically themed prompts for several non-Western contexts rather than distributed randomly.DALL-E-3 showed especially high refusal rates for Chinese Simplified–China, Chinese Traditional–Taiwan, Ukrainian–Ukraine, and Russian–Russia compared with Anglophone contexts.
  • Imagen: Imagen automatically revised prompts and returned them with images, generating one 1024×1024 image per prompt because of budget constraints.Imagen successfully generated 8,960 images across combinations before exclusions for refusals in the broader procedure.
  • GPT-Image: GPT-Image allowed separate specification of the revision and image-generation models; the study used gpt-5.4-nano, 1024×1024 low-quality images, and standardized prompt prefixes.These choices were made for budget and input-framing consistency.
  • Guardrails: Main Results analyses use only cases where an image and revised prompt were successfully generated, while refusal patterns are documented separately as culturally asymmetric guardrail behavior.Figure 5 organizes refusal rates by model, category, and language–context combination.

C GPT Image non-English revised prompts and backtranslation

GPT-Image sometimes returned revised prompts in their source language, requiring backtranslation for comparison, while the analysis filtered geographic and linguistic artifacts before measuring cultural terms.

  • C GPT Image non-English revised prompts and backtranslation: Unlike DALL-E-3 and Imagen, GPT-Image sometimes preserved the original prompt language during revision instead of always translating to English.DALL-E-3 and Imagen returned all revised prompts in English and exposed no user control over this processing step.
  • C GPT Image non-English revised prompts and backtranslation: Table 2 provides illustrative original-versus-revised prompt examples across language–context conditions.These examples complement the quantitative analysis of revision behavior.
  • C GPT Image non-English revised prompts and backtranslation: GPT-Image returned some revised prompts in non-English languages, with shares ranging from 0% to 100% across language–context combinations and prompt categories.The variation is summarized in Figure 6.
  • C GPT Image non-English revised prompts and backtranslation: Non-English revised prompts were language-detected, translated back into English, and manually quality-checked before cross-language comparison.Lingua was used for detection; Tower-Plus-9B handled most translations and Qwen2.5-7B-Instruct handled Arabic.
  • D Geographic and Artifact Stopwords: The TF-IDF analysis removes country, city, landmark, language, script, translation, VLM, filler, and function-word artifacts so distinctive terms better reflect cultural injection.Without filtering, geographic identifiers such as country names would dominate context-specific term lists.
  • D Geographic and Artifact Stopwords: Generic evaluative padding, foreign-language function words, and residual translation fragments are filtered because they lack reliable cultural signal.Examples include “breathtaking,” “charming,” language function words, and instruction leakage from imperfect translation.
  • D Geographic and Artifact Stopwords: The stopword list excludes regional labels such as “Nordic” and “Mediterranean” because the authors treat them as culturally interpretive signals.This choice distinguishes geographic tautology from potentially meaningful cultural categorization.

E CFS Robustness

CFS rankings remain stable across hyperparameter settings, while text-level CMS is positively related to image-level CMS across all three models.

  • CFS Robustness: CFS rankings are stable across configurations, with Kendall’s τ ranging from 0.77 to 0.91 within quantiles and 0.77 to 0.84 across quantiles.The TF-IDF threshold is the most sensitive setting; the 50th percentile diverges most from the default.

I Visual-Level Analysis: Full Results

The visual-level analysis compares textual and visual cultural flattening across models and contexts, while testing how revised-prompt terms propagate into generated images. Textual and visual flattening are positively related, but the propagation analysis remains correlational.

  • Visual-Level Analysis: Full Results: The visual-level extension applies the revised-prompt TF-IDF and CFS framework to VQA descriptions across all three commercial models and 31 contexts.The analysis extends the same framework used for revised prompts to visual descriptions.
  • Visual-Level Analysis: Full Results: Contexts with high text-level flattening tend to show high visual-level flattening, although the relationship varies by model.Figure 8 covers all contexts and models; the supplied materials do not report correlation values.
  • Visual-Level Analysis: Full Results: Visual distinctive terms are compared with revised-prompt terms to classify which terms are propagated and which are image-model only.Propagation rates are summarized across models in Table 6, with nmaj identifying contexts where more than half of visual terms are propagated.
  • Visual-Level Analysis: Full Results: The propagation analysis is correlational because revised terms may appear visually through the revision layer or independently through the image model.The ablation study is intended to address this attribution limitation.

J.1.2 Metrics

The matched-pair ablation compares images from original and revised prompts to measure revision effects on cultural markedness, flattening, and distinctive terms. Across two open-weight models, revisions substantially shift cultural content and amplify stereotypical markers, though causal evidence beyond English-speaking contexts is limited.

  • J.1.2 Metrics: Flux images receive descriptions seven words longer on the median, whereas SDXL shows a nonsignificant one-word difference.For Flux, V = 560,627 and p < 0.001; for SDXL, p = .12.
  • J.1.2 Metrics: All four Flux contexts and three of four SDXL contexts are statistically significant, with India+English nonsignificant for SDXL.The SDXL India+English result is p = .09, consistent with culturally marked content already appearing from original prompts.
  • J.1.2 Metrics: CFS increases for every context–model combination, with the largest increases in the UK and the smallest in the US and India.The US receives little distinctive vocabulary, whereas India is near a CFS ceiling in both conditions.
  • J.1.2 Metrics: Across both models, 13–15 terms are significantly amplified and 1–2 are suppressed, with amplified terms functioning as stereotypical cultural markers.For India, suppressed terms such as doorway and unpaved suggest replacement of generic visual markers with culturally specific ones.
  • J.1.2 Metrics: Revised-only terms form coherent stereotypes, while original-only terms are generally less stereotypical except for independently generated markers in India and partly the UK.Examples include outback and nature for Australia, traditional dress and festivals for India, village life and London imagery for the UK, and suburban Americana for the US.
  • J.1.2 Metrics: The causal ablation covers four English-speaking contexts, leaving non-English high-flattening contexts largely untested.A Switzerland extension addresses one non-English case, but only one such context was examined and language effects may be confounded by inadequate prompt representation.
  • J.1.2 Metrics: Median cosine distance from the unrevised baseline is 0.78 for SDXL and 0.69 for Flux after prompt revision.The revision effect introduces recognizably cultural terms such as lederhosen, fondue, and watchtower rather than generic vocabulary.
Loading 2609.11532v1…