You Asked for an "Egyptian Living Room" — the AI Quietly Added Pyramids for You

Imagine opening DALL-E and typing "a living room" (in Arabic, with "Egypt" as context). You expect a sofa, coffee table, and TV from a Cairo home. Instead you get pyramids, a sphinx, and hieroglyphic decor.

You might think this is the image model "misunderstanding." It isn't. The problem lies in an invisible layer between you and the image model — the prompt revision layer. After you hit enter and before generation, the system quietly rewrites "a living room" into "an Egyptian living room with pyramids and pharaonic decorations."

The image model faithfully draws whatever instructions it receives. The real "seasoning" comes from the ghost in the middle.

A recent paper did something no prior work had: it split open the T2I pipeline and audited the hidden revision layer in isolation. The results are striking.

Key points

• The WORLDVIEW benchmark audits the prompt revision layer of DALL-E-3, Imagen-4, and GPT-Image-1.5 using 8,960 prompts: 15 languages × 31 language-culture pairs across 14 domains (home, transport, politics, religion, work, holidays, etc.), each with an unmarked English baseline ("a living room") and a contextualized version ("a living room in Egypt," in Arabic).

Finding 1: Contextual Marking (CMS) — America is the "unmarked default"

The Contextual Marking Score measures how much the revision layer changes culturally-contextualized prompts relative to unmarked English baselines. A clear marking ladder emerges:

| Context | CMS range | Meaning | |---------|-----------|---------| | USA | 0.21–0.31 | Almost untouched — America is the default | | UK | 0.24–0.35 | Slight changes — the Anglosphere is "near-default" | | Germany | Moderate | European core | | Finland | Highest (up to 0.47) | Most modification — furthest from the Anglo default | | MENA | High | Arabic contexts heavily "supplemented" |

Systems add almost nothing to "an American living room" because they assume a living room *is* American by default. Egypt's living room, meanwhile, gets layers of "cultural" additions.

Finding 2: Cultural Flattening (CFS) — a few keywords stand in for a whole culture

The Cultural Flattening Score combines token concentration (how many inserted words are culture-specific) with token diffusion (how many unrelated prompts they invade):

• Switzerland: highest CFS. "alps," "chalet," "chocolate," "fondue" appear in 77–96% of prompts — living rooms, political rallies, and ads alike.
• Finland: "snow," "pine," "northern," "birch" in nearly everything.
• Saudi Arabia: "desert," "thobe," "palm," "islamic" everywhere.
• Egypt: "pyramid," "sphinx," "pharaonic," "hieroglyph" — contemporary Egypt compressed into ancient Egypt.
• Mexico: "sombrero," "mariachi," "cactus," "tacos."

CMS and CFS are distinct: Switzerland gets moderate CMS but the highest CFS — fewer edits, but the most indiscriminate stereotyping. Arabic contexts receive many edits (high CMS) that are relatively varied (moderate CFS).

Finding 3: Causal attribution — the revision layer is the *cause*, not a passenger

The paper feeds both original and revised prompts into revision-free open-source models (SDXL Lightning and Flux-2-Dev):

• Images from revised prompts show significantly higher CMS (Wilcoxon, p < 0.001).
• CFS rises from 0.75→0.86 (SDXL) and 0.85→0.98 (Flux); the UK shows the largest CMS jump (+34% SDXL, +52% Flux).
• Revised-prompt-specific tokens are recognizable stereotypes: "outback," "kangaroo" (Australia); "cobblestone," "pub" (UK); "marigold," "taj" (India).
• India is a partial exception: image models themselves generate "bindi," "dhoti," "curry" even from raw prompts, indicating training-data bias — but the revision layer adds a *second* stereotype set ("marigold," "rangoli," "taj").
• Conclusion: 33–46% of visual stereotype tokens are injected at the revision stage. The image model faithfully executes the revision layer's instructions.

Why this matters

The audit blind spot. Prior T2I bias research treats the whole pipeline as a black box, comparing only input to output. WORLDVIEW's contribution is decomposing the pipeline and auditing each stage — revealing that the main source of bias is not the image model but the neglected revision layer. Fixing the image model is like replacing a door lock while the key sits with the revision layer.

The politics of defaults. The deepest finding is not that Egypt gets pyramids, but that America gets nothing. The US is the unmarked default — like "skin color" labels where white is unmarked and everything else is "colored." This reflects a Western-centric training corpus: non-Western contexts are treated as "anomalies requiring supplementation."

Repair implications. A faithful image model still produces stereotyped images if its input has already been stereotyped. Interventions must target the revision layer — making it culturally sensitive rather than reducing a country to five keywords, and letting users see, review, and reject specific edits. The authors note that even this is insufficient: training data, architecture, and evaluation paradigms need parallel fixes.

Broader methodological takeaways

The three-step audit (CMS → CFS → stereotype content analysis, plus causal ablation) generalizes beyond cultural bias — to gender, race, or occupational bias in any pipeline. And the "ghost layer" phenomenon is universal: search engine query rewriting, recommender user-profile expansion, translation "polish," and coding-assistant style correction are all invisible, non-disableable, unauditable middle layers whose definition of "better" is set by the system. Transparency and user control over these layers may be a more direct remedy than fixing models alone.

---

Paper: Urman, A., Lichtenegger, E., Hannak, A., et al. *WORLDVIEW: Auditing the Prompt Revision Layer in Commercial Text-to-Image Systems*. arXiv:2609.11532.

Code: github.com/aurman21/worldview_prompt-revision

Data: 8,960 prompts × 3 systems, covering 15 languages, 31 cultural contexts, and 14 domains.