Cultural Ghosting: AI & Cultural Signal Loss

• Cultural ghosting is the phenomenon where AI systems erase or simplify culturally specific markers, leading to generic and dominant representations.
• Quantitative studies reveal that conventional metrics like CLIPScore and SPS often overlook cultural fidelity losses, highlighting the need for specialized evaluation methods.
• Mitigation strategies such as explicit cultural prompts, fine-tuning interventions, and neuron-level enhancements show promise in preserving cultural integrity in AI outputs.

Cultural ghosting is a cross-domain term for the fading, erasure, or flattening of culturally specific signals in AI-mediated representation. In generative image models, it denotes the systematic distortion or erasure of underrepresented cultures and the replacement of culturally meaningful detail with dominant defaults or token surface cues (Seo et al., 22 Oct 2025). In multilingual text-to-image generation, it names the tendency for non-English prompts to yield culturally neutral or subtly English-biased images unless explicit cultural modifiers are supplied (Shi et al., 21 Nov 2025). In LLMs, the term has been used for the erasure of linguistic markers in World Englishes during rewriting, for reversion toward restricted default value profiles in value surveys, and for omission or simplification of cultural representation in knowledge and recommendation tasks (Navneet et al., 25 Feb 2026, Luther et al., 13 Dec 2025, Bulté et al., 6 Nov 2025, Qadri et al., 2 Jan 2025, Tao et al., 2023). A distinct CSCW usage applies the term to users’ selective withdrawal from culturally misaligned digital content through pragmatic disengagement and interdependent navigation (Seo et al., 23 May 2025).

1. Conceptual scope and terminological variation

The literature does not use cultural ghosting as a single, fully stabilized term. Instead, several adjacent usages converge on a common structure: culturally salient information is present, requested, or socially consequential, yet outputs or interactions suppress it, flatten it, or route it toward a dominant norm.

Domain	Meaning of cultural ghosting	Representative source
Generative images	Erasure or superficial masking of target culture	(Seo et al., 22 Oct 2025)
Multilingual T2I	Culture-specific signals fade under noun-only prompting	(Shi et al., 21 Nov 2025)
LLM rewriting	World-English markers are stripped during editing	(Navneet et al., 25 Feb 2026)
LLM cultural values	Outputs revert to restricted default national profiles	(Luther et al., 13 Dec 2025, Bulté et al., 6 Nov 2025, Tao et al., 2023)
CSCW non-use	Users selectively withdraw from misaligned digital spaces	(Seo et al., 23 May 2025)

A closely related framing comes from work on cultural erasure in LLMs, which distinguishes omission from simplification. Omission concerns whether a culture appears at all, whereas simplification concerns how a culture is represented when it does appear. This distinction supplies a sociologically aware vocabulary for understanding ghosting as either absence or one-dimensional reduction (Qadri et al., 2 Jan 2025).

The image-model literature adds a more processual interpretation. Seo et al. describe ghosting not only as a default to Global-North, modern-leaning depictions under country-agnostic prompts, but also as a cumulative effect in iterative editing, where repeated edits progressively strip away meaningful cultural detail and leave only palette shifts, flags, or generic props (Seo et al., 22 Oct 2025). By contrast, the CSCW study of older Korean immigrants treats ghosting as a user practice: elders selectively “ghost” content and features that are linguistically opaque, affectively taxing, or culturally misaligned (Seo et al., 23 May 2025). This suggests that the term now spans both model-side erasure and user-side withdrawal.

2. Formalization and evaluation frameworks

Several papers make cultural ghosting measurable through explicit metrics. In the cultural-erasure framework for LLMs, thematic presence for a city $c$ under theme $t$ is defined as

$P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$

which aggregates to a regional thematic score

$S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$

In travel recommendation, regional omission is measured by

$$P_{rec}(R)=\frac{1}{M}\sum_{m=1}^{M}\mathbf{1}_{f(c_m)=R}.$$

Regions with $P_{rec}(R)\approx 0$ are effectively omitted (Qadri et al., 2 Jan 2025).

The unified T2I/I2I benchmark for cultural bias in generative image models defines a cluster-proportion representation over countries and models,

$$p^{(m)}_{c,k}=\frac{\#(i:c(i)=c\wedge k(i)=k)}{\#(i:c(i)=c)},$$

then combines cosine similarity and Jensen–Shannon divergence into a proximity score $h^{(m)}_{ab}$ for country pairs. The same study also defines a traditional–modern leaning score

$$s_i=\cos(x_i,\mu_{trad}(g(i)))-\cos(x_i,\mu_{mod}(g(i))),$$

with country-level means and standard errors derived from $s_i$. Automatic evaluation uses CLIPScore, DreamSim distance, and Aesthetic Score, while culture-aware evaluation uses retrieval-augmented VQA and human evaluation uses a 1–5 Likert protocol collapsed into a Human Quality Score (HQS) (Seo et al., 22 Oct 2025).

In multilingual T2I, cultural consistency is quantified by CultureVQA accuracy,

$t$0

where the VQA model chooses one of 15 region labels or “unrecognisable.” Cultural ghosting is then measured by the drop from “culture-style modifier + noun” prompts to noun-only prompts,

$t$1

A large gap, often 20–30 points, indicates strong ghosting (Shi et al., 21 Nov 2025).

In LLM rewriting, Navneet et al. introduce two complementary metrics. Identity Erasure Rate is

$t$2

when $t$3, and Semantic Preservation Score is cosine similarity between sentence embeddings,

$t$4

This design isolates semantic fidelity from stylistic or cultural change (Navneet et al., 25 Feb 2026).

For cultural values in LLMs, two survey-based families of metrics recur. Hofstede-based work computes Euclidean distance between a model-prompt six-vector and a country’s Hofstede vector, while IVS-based work projects model responses into the Inglehart–Welzel cultural map and measures Euclidean distance there. In both cases, smaller distance implies higher cultural alignment (Luther et al., 13 Dec 2025, Tao et al., 2023).

3. Image generation and editing as sites of ghosting

Research on text-to-image and image-to-image systems has progressively moved from prompt-sensitivity studies to standardized audits. An early TTI study proposed a three-tier cultural ontology spanning cultural dimensions, cultural domains, and cultural concepts; prompt templates such as fully translated prompts, translated concepts, and “English with Nation”; and intrinsic metrics in CLIP space including National Association, Cultural Dimensions Projection, Cultural Distance, and Cross-Cultural Similarity. Across four TTI models and ten languages, the strongest cultural unlock was explicit nationality naming in English, while mere translation or gibberish cues were weaker. The same work described cultural ghosting as the production of generic images lacking salient culture-specific visual cues, with case studies including “Indian wedding” and “city” prompts that frequently reverted to Western defaults unless stronger cues were provided (Ventura et al., 2023).

Seo et al. extend this line with a unified evaluation across six countries—China, India, Kenya, Korea, Nigeria, and the U.S.—using an 8-category/36-subcategory schema and three era-aware prompts per $t$5: traditional, modern, and era-agnostic. Experiments cover five open-source model families in T2I and I2I modes, and three I2I paradigms: Multi-Loop Edit, Attribute Addition, and Cross-Country Restylization. Human evaluation is performed by 17 country-native experts, each rating only their own country, with side-by-side review of base, step 1, step 3, and step 5 images (Seo et al., 22 Oct 2025).

The principal empirical result is a consistent Global-North default. Under country-agnostic prompts, all five T2I models cluster most closely with U.S. prompts, with $t$6 and 95% CI $t$7. China–Korea and Kenya–Nigeria form strong regional pairs with $t$8 and $t$9. Traditional–modern leaning under era-agnostic prompts shows a consistent modern bias for U.S. and agnostic cases, with $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$0, $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$1, and FDR $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$2. Qualitatively, FLUX.1 and HiDream produce nearly identical “Bride and groom,” “Chef,” and “Farmer” scenes whether prompted “in U.S.” or country-agnostically (Seo et al., 22 Oct 2025).

The same benchmark identifies a strong dissociation between conventional metrics and cultural fidelity in iterative editing. From step 0 to step 5, CLIPScore remains flat or rises modestly, with $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$3mean $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$4 and range $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$5, while HQS declines by an average of 44.2%, often from about 4 at base to about 1–2 by step 5. The culture-aware VQA metric tracks HQS, with Best/Worst agreement of 73.8% and 83.7%. In attribute addition, local-script text frequently degenerates into gibberish and the final accessory step yields the lowest image quality; in cross-country restylization, models often change only palette, flags, symbols, or painterly style while retaining U.S.-like subject identity for Global-South targets (Seo et al., 22 Oct 2025).

A complementary multilingual T2I study argues that ghosting may reflect under-activation rather than absent knowledge. By probing attention contrast and then using a Top-K sparse autoencoder, it localizes culture-sensitive signals to a small set of neurons in a single “culture layer,” finding layer 16 in PEA-Diffusion and layer 14 in AltDiffusion. On CultureBench, inference-time cultural activation and layer-targeted cultural enhancement improve noun-only prompt performance from 21.65 to 33.91 and 36.63 CultureVQA in PEA-Diffusion, and from 23.05 to 30.06 and 32.66 in AltDiffusion, while maintaining or improving CLIPScore, ImageReward, and LPIPS. An ablation in PEA-Diffusion shows that masking Top-K neurons collapses CultureVQA from 35.62 to 7.65, whereas random masking yields 33.04 (Shi et al., 21 Nov 2025).

4. LLMs, cultural values, and linguistic voice

In LLM-mediated writing, cultural ghosting has been operationalized as the erasure of linguistic identity during “professionalization.” Navneet et al. analyze 22,350 outputs generated from 1,490 culturally marked texts in Indian, Singaporean, and Nigerian English, processed by five open-source instruction-tuned models under Baseline, Neutral, and Preservation prompts. Across all outputs, mean IER is 10.26%, model-level IER ranges from 3.5% to 20.5%, and mean SPS is 0.748. Pragmatic markers are 1.9x more vulnerable than lexical markers, with 71.5% versus 37.1% erasure. The Preservation prompt reduces IER by 29% relative to Baseline, from 0.116 to 0.082, while slightly increasing SPS from 0.742 to 0.755. The paper terms this pattern the Semantic Preservation Paradox: models preserve meaning while systematically erasing cultural markers (Navneet et al., 25 Feb 2026).

A second line of work studies cultural ghosting at the level of value profiles. Luther and Brown use Hofstede’s six dimensions and the VSM 2013 International Survey’s 24 questions to assess eight flagship LLMs. Without cultural prompting, five models align most strongly to the U.S.; two default closest to Iran; and Mistral Large shows no “soft” or “strong” alignment to any tested country. Cultural prompting reduces cumulative distance for seven of eight models, but models remain farthest from China’s real-world Hofstede values and also struggle with Japan. The paper names this persistent default the “American ghost” (Luther et al., 13 Dec 2025).

A broader multilingual values study probes 10 LLMs with 63 items from Hofstede’s Values Survey Module and the World Values Survey in 11 languages and four prompt variants. It finds that prompt language and cultural framing both induce variation, but models remain systematically biased toward the Netherlands, Germany, the United States, and Japan. Averaged over all 23 WVS items and ten models, Pearson $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$6 with human country means is .75 for Germany, .72 for the United States, .71 for the Netherlands, and .68 for Japan, versus about .28–.47 for the other seven countries. Explicit cultural perspective improves alignment more than prompt language alone, yet $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$7 and $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$8 remain nearly identical at $P_t(c)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}_t(d_i),$9 and $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$0, and combining language with perspective does not outperform English-only cultural framing (Bulté et al., 6 Nov 2025).

An earlier IVS-based analysis of five GPT models reaches a similar conclusion from a different survey geometry. Default model outputs cluster near English-speaking and Protestant European countries on the Inglehart–Welzel map and are farthest from African-Islamic countries. Cultural prompting reduces average cultural distance for all models and improves 71.0% to 81.3% of countries, depending on model, with an extreme case in which GPT-4o’s distance to Jordan drops from 4.10 to 0.36. However, some Western countries move farther away under prompting, including Finland from 0.20 to 2.43 and Luxembourg from 0.59 to 2.72 (Tao et al., 2023). Taken together, these studies support the view that cultural prompting can steer outputs, but only within a landscape still anchored to restricted defaults.

5. Omission, simplification, and user-side withdrawal

The omission/simplification framework broadens cultural ghosting beyond direct prompting tasks. In place-description probes using PaLM, 50 major global cities are described under five open-ended templates, producing 2,500 descriptions annotated for themes such as culture and economy. European and North American cities obtain the highest $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$1, while African and Asian cities are more likely to evoke economic themes. For Africa, $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$2 and $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$3, with pairwise $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$4-tests showing Africa’s culture-theme score significantly lower than Europe’s and its economy-theme score significantly higher. In travel recommendation with 280 samples, eight subregions—including all of Central, Eastern, and Western Africa—have $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$5, while Western Europe and Eastern Asia dominate with approximately 0.411 and 0.407 (Qadri et al., 2 Jan 2025).

The CSCW literature uses the same term in a different but related way. A CBPR study of older Korean immigrants in the greater NYC area identifies pragmatic disengagement and interdependent navigation as two key practices through which elders manage digital noise, language exclusion, and emotional fatigue. Twenty-two participants aged 65–85 completed 45–60 minute semi-structured interviews in Korean, recruited through a Korean grocery store in New Jersey, Borisa Buddhist temple in Englewood, and a senior daycare center in Nutley. Analysis proceeded through familiarization, open coding, code consolidation, theme development, cross-case comparison, and cultural validation, with three community advisors reviewing preliminary themes (Seo et al., 23 May 2025).

Here, cultural ghosting denotes selective disappearance from misaligned parts of digital systems rather than representational failure by models. Participants report turning off combative political videos, abandoning app installation when English-only login prompts appear, avoiding online banking because it feels too risky, or leaving KakaoTalk group chats after mishaps. The study frames these behaviors not as deficit-oriented non-use but as thoughtful, protective, and culturally situated action. It extends CSCW theories of non-use, layered uncertainty, and folk algorithmic literacy by treating ghosting as bottom-up algorithmic resistance and as a means of preserving emotional safety and cultural coherence (Seo et al., 23 May 2025).

6. Mitigation strategies and unresolved problems

A recurring result across modalities is that conventional quality or semantic metrics do not reliably track cultural fidelity. In I2I editing, CLIPScore can stay flat or improve while HQS collapses; in LLM rewriting, SPS remains high even when cultural markers are erased; and in multilingual values studies, outputs vary under prompting without escaping restricted default profiles (Seo et al., 22 Oct 2025, Navneet et al., 25 Feb 2026, Bulté et al., 6 Nov 2025). A plausible implication is that culture-sensitive evaluation must be treated as a first-class objective rather than as a by-product of alignment, realism, or semantic preservation.

The image-model literature proposes several concrete responses. Seo et al. release a standardized benchmark comprising images, prompts, model settings, and evaluation code to support diagnosis and mitigation, including balanced, subnational, era-diverse training sets, era-preservation and culture-consistency losses, and regularization that penalizes superficial cues such as flags and palettes when they substitute for deeper cultural content (Seo et al., 22 Oct 2025). The multilingual T2I probing work adds two lightweight interventions: a “Zero-Training Neuron Amplifier,” which amplifies identified culture-sensitive neurons by a factor $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$6, and a “Fine-Tuned Layer Enhancer,” which inserts a residual adapter into the detected culture layer while freezing the backbone (Shi et al., 21 Nov 2025).

For LLM rewriting, mitigation spans prompt engineering and decoding. Explicit preservation prompts reduce IER by 29% without semantic trade-off, while marker-aware constrained decoding drops IER by 47% on a 200-text subset and contrastive reranking with $S_t(R)=\frac{1}{|C|}\sum_{c\in C}P_t(c).$7 achieves a 31% reduction (Navneet et al., 25 Feb 2026). For value alignment and broader representational coverage, recommended interventions include diversify pretraining data, fine-tuning on structured value-survey data stratified by culture and language, explicit cultural metadata or latent culture embeddings, anthropological prompting, post-hoc calibration layers, multicultural RLHF, transparent reporting of cultural alignment, and ongoing evaluation with benchmarks such as CDEval and Hofstede-CAT (Luther et al., 13 Dec 2025, Bulté et al., 6 Nov 2025, Tao et al., 2023).

More general recommendations come from the cultural-erasure framework: multidisciplinary benchmarks, rich annotation schemes, data curation and augmentation using under-represented narratives, modeling interventions that enforce cultural-theme coverage thresholds, and collaboration with cultural domain experts and affected communities (Qadri et al., 2 Jan 2025). The CSCW study proposes bilingual “Core Mode,” customizable curation filters, joint-use links for trusted helpers, “Ghost Trackers,” tech-mediation carework, community tech liaisons, and interest-anchored literacy modules (Seo et al., 23 May 2025). Across these papers, the central unresolved problem is not merely whether models can represent culture, but whether they can do so without defaulting to omission, simplification, semantic-only preservation, or dominant cultural priors.