💬 NLP

When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models

This study introduces the concept of "Cultural Ghosting" to quantify how Large Language Models systematically erase linguistic markers from non-native English varieties like Indian, Singaporean, and Nigerian English, revealing a paradox where semantic meaning is preserved while cultural identity is lost, though explicit prompts can significantly mitigate this erasure.

Original authors: Satyam Kumar Navneet, Joydeep Chandra, Yong Zhang

Published 2026-02-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Satyam Kumar Navneet, Joydeep Chandra, Yong Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Cultural Ghosting" of AI

Imagine you have a friend who writes emails in a very specific, warm, and culturally rich way. They use phrases like "Kindly do the needful" or "Respected Sir." These aren't just words; they are like cultural spices that add flavor, respect, and hierarchy to the message.

Now, imagine you ask an AI assistant to "make this email more professional." The AI comes back and says, "Please complete the task" and "Hello."

The meaning is exactly the same. The task is clear. But something vital is missing. The "spices" are gone. The unique voice of the writer has been scrubbed clean, leaving behind a bland, generic version of English that sounds like it was written by a robot from a specific place (usually the US or UK).

The researchers call this "Cultural Ghosting." It's like the AI is a ghost that haunts your writing, quietly stealing your cultural identity while pretending to just "help" you.

The Experiment: Testing the Ghosts

The researchers wanted to see how bad this problem is. They didn't just guess; they ran a massive experiment.

The Ingredients: They took 1,490 real texts written in three specific types of English: Indian, Singaporean, and Nigerian. These are known as "World Englishes"—variations of English that have their own unique rules, words, and politeness styles.
The Chefs: They fed these texts into 5 different AI models (the "chefs").
The Orders: They gave the AI three different instructions:
1. "Make this professional." (The standard order)
2. "Make this clear." (The neutral order)
3. "Make this clear, but keep the cultural voice." (The preservation order)
The Result: This created 22,350 different rewritten emails.

They then measured two things:

Did the AI keep the meaning? (Yes, usually very well).
Did the AI delete the cultural "spices"? (Yes, often very aggressively).

The Big Surprise: The "Semantic Preservation Paradox"

Here is the weird part that the researchers discovered. They call it the Semantic Preservation Paradox.

Think of it like a photocopier.

If you copy a colorful, hand-drawn map, a bad photocopier might turn it black and white. You can still read the street names (the meaning is preserved), but the colors and artistic style (the culture) are gone.
The AI is doing exactly this. It is keeping the "street names" (the message) perfectly clear, but it is turning the "colors" (the cultural identity) into gray.

The Numbers:

The AI kept the meaning about 75% of the time (which is great).
But it erased about 10% of the cultural markers on average.
In the worst AI models, they erased up to 20% of the cultural voice!

Who Gets Hurt the Most?

The researchers found that not all "spices" are treated equally.

The "Politeness" Spices (Pragmatic Markers): These are the most vulnerable. Words like "Kindly," "Respected Sir," or particles like "Lah" (used in Singapore) are erased 71.5% of the time.
- Analogy: Imagine a chef who is told to "fix the soup." They remove the salt and the pepper because they think it's "too much flavor," not realizing that the salt and pepper are the flavor. The AI removes the politeness because it thinks "directness" is more professional.
The "Vocabulary" Spices (Lexical Markers): Words like "Do the needful" or "Prepone" are erased less often (about 37%).
- Analogy: The AI is more willing to keep a weird ingredient if it's a noun, but it immediately deletes the "seasoning" words that show respect or relationship.

The Takeaway: The AI is most dangerous to the parts of language that show respect, hierarchy, and community. It flattens these relationships into cold, direct commands.

Can We Stop the Ghost?

Yes! The researchers found two ways to stop the erasure without making the writing sound bad.

1. The "Magic Spell" (Prompt Engineering):
If you simply tell the AI, "Please improve this, but keep the original cultural voice and phrasing," the erasure drops by 29%.

Analogy: It's like telling the photocopier, "Make this clearer, but do not turn off the color." The AI listens, and suddenly, the "Kindly" and "Respected Sir" stay in the email.

2. The "Safety Net" (Algorithmic Fixes):
The researchers also tried a technical fix where they forced the AI to "copy" certain words if they were cultural markers. This reduced erasure by 47%.

Why Does This Matter?

The paper argues that this isn't just about grammar; it's about identity.

For billions of people who speak "World Englishes," their way of speaking is not "broken" English that needs fixing. It is a valid, rich way of communicating that carries social meaning. When AI automatically "corrects" this to sound like American or British English, it is essentially whitewashing their identity.

The Final Lesson:
The fear that "we have to choose between being clear and being ourselves" is a myth. You can have both. The AI is currently designed to erase culture by default, but with simple changes (like telling it to keep your voice), we can stop the "ghosting" and let our unique cultural voices remain in the digital world.

In short: AI is currently a "cultural eraser." But if we tell it to be a "cultural preserver," it can keep our unique voices alive while still helping us write clearly.

1. Problem Statement: Cultural Ghosting

The paper addresses a critical, under-explored bias in Large Language Models (LLMs): Cultural Ghosting. This phenomenon occurs when LLMs, tasked with "professionalizing" or polishing text, systematically erase linguistic markers unique to non-native English varieties (World Englishes) while preserving the core semantic meaning.

The Core Issue: Unlike explicit toxicity or stereotyping, this erasure happens during benign tasks like grammar correction and email polishing.
The Semantic Preservation Paradox: The authors identify a tension where models achieve high semantic similarity (preserving what is said) while simultaneously erasing cultural identity (removing how it is said).
Impact: For the 1.5+ billion speakers of World Englishes (e.g., Indian, Singaporean, Nigerian English), this results in "cultural whitewashing," stripping away social context, pragmatic functions (like hierarchical politeness), and cultural identity.

2. Methodology

The study employs a rigorous quantitative approach to measure and analyze this phenomenon across 22,350 generated outputs.

A. Dataset Construction

Corpus: 1,490 culturally marked texts were curated from emails (Enron, Clinton Archive), social media, and news.
Varieties: The dataset includes Indian English ( $n=601$ ), Singaporean English ( $n=261$ ), Nigerian English ( $n=89$ ), and an American English baseline ( $n=539$ ).
Annotation: A lexicon of 108 culturally specific markers was compiled, categorized into:
- Lexical: Unique vocabulary (e.g., "prepone," "chope").
- Pragmatic: Politeness and social positioning (e.g., "kindly," "Respected Sir," "lah").
- Syntactic: Grammatical structures (e.g., "discuss about").
Validation: Automated annotation was validated with high inter-annotator agreement ( $\kappa=0.89$ ).

B. Experimental Setup

Models: Five open-source instruction-tuned LLMs were evaluated: Mistral-7B, Apertus-8B, DeepHat-7B, MiMo-7B, and Qwen3-8B.
Prompt Conditions: Each text was processed under three conditions:
1. Baseline: "Make this text more professional & grammatically correct."
2. Neutral: "Improve the clarity & grammar of this text."
3. Preservation: "Improve clarity & grammar while preserving cultural voice & regional expressions."

C. Evaluation Metrics

The authors introduced two novel metrics:

Identity Erasure Rate (IER): Quantifies the proportion of cultural markers removed.
$IER = \frac{M_{original} - M_{output}}{M_{original}}$
(Ranges from 0 to 1, where 1 is complete erasure).
Semantic Preservation Score (SPS): Measures cosine similarity between sentence embeddings to ensure meaning is retained.
$SPS = \frac{e_{orig} \cdot e_{out}}{\|e_{orig}\|\|e_{out}\|}$
(Validated against human judgment with Pearson $r=0.82$ ).

3. Key Contributions

Formalization of Cultural Ghosting: Defined the phenomenon and the "Semantic Preservation Paradox" where high semantic fidelity coexists with identity loss.
Novel Metrics: Introduced and validated IER and SPS, providing the first large-scale quantification of marker erasure.
Empirical Evidence of Vulnerability: Demonstrated that pragmatic markers are significantly more vulnerable than lexical ones.
Mitigation Strategies: Proved that explicit prompts and algorithmic constraints can reduce erasure without sacrificing semantic quality.

4. Key Results

A. Extent of Erasure & Model Variation

Overall Erasure: The mean IER across all models and prompts was 10.26%.
Model Disparity: Erasure rates varied significantly by model (from 3.5% for Qwen3-8B to 20.5% for Mistral-7B), a 5.9× range.
Drivers: The variation was driven by alignment strategies (specifically RLHF training data diversity) rather than model parameter count. Models with intensive RLHF showed divergent behaviors based on their training corpus composition.

B. Differential Marker Vulnerability

Pragmatic vs. Lexical: Pragmatic markers (politeness, honorifics) were 1.9× more likely to be erased than lexical markers.
- Pragmatic Erasure: 71.5%
- Syntactic Erasure: 56.3%
- Lexical Erasure: 37.1%
Reasoning: Pragmatic markers encode "face-management" and social hierarchy that are not recoverable from semantic content alone, making them the primary target of "standardization" optimization.

C. Mitigation via Prompts and Algorithms

Prompt Engineering: Using a "Preservation" prompt reduced the IER by 29% compared to the baseline, with no loss in Semantic Preservation Score (SPS actually improved slightly).
Algorithmic Mitigation:
- Marker-aware constrained decoding: Reduced IER by 47% on a subset.
- Contrastive reranking: Reduced IER by 31%.
The Trade-off Myth: Regression analysis ( $R^2 = 0.061$ ) confirmed that cultural authenticity and semantic clarity are largely independent. Preserving culture does not degrade clarity.

5. Significance and Implications

Design Choice, Not Inevitability: The paper argues that cultural ghosting is a result of current alignment paradigms (favoring Western norms) rather than a technical limitation. It is a design choice that can be reversed.
Workplace Impact: In professional settings, erasing markers like "Kindly" or "Respected Sir" removes crucial institutional hierarchy and relational context, potentially altering the power dynamics of communication.
Recommendations for Developers:
1. Default to Preservation: Shift default settings to preserve cultural voice rather than forcing convergence.
2. Transparency: Provide users with "keep my phrasing" options.
3. Variety Recognition: Adapt models to recognize and respect World Englishes rather than treating them as errors to be corrected.
4. Training Adjustments: Implement variety-specific fine-tuning and politeness-aware reward modeling.

Conclusion: The study concludes that while current LLMs act as "cultural standardization engines," the trade-off between identity and clarity is a myth. Through prompt engineering and algorithmic constraints, it is possible to maintain high semantic quality while retaining the unique cultural voices of World English speakers.