Elamur Research⌕
☰
home / papers / 2026 / When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models

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

Paper · 2026 · Satyam Kumar Navneet, Joydeep Chandra and Yong Zhang

Summary

The paper investigates "cultural ghosting": when people ask an LLM to make text more professional or clearer, the model quietly removes features typical of non-native English varieties while keeping the meaning. The authors built a corpus of 1,490 texts drawn from email collections (Enron, Clinton Archive, EmailSum), Twitter/X posts and news. The texts cover Indian, Singaporean and Nigerian English, plus an American English baseline. Each text was rewritten by five open instruction-tuned models under three prompts, giving 22,350 outputs. Two metrics score each rewrite. Identity Erasure Rate (IER) is the share of lexicon markers that disappear. Semantic Preservation Score (SPS) is the embedding cosine similarity between input and output. Results:

  • Meaning is largely kept (SPS about 0.75), while about a tenth of markers vanish. The authors call this the Semantic Preservation Paradox.
  • Politeness-type markers are the most vulnerable.
  • Asking the model to keep cultural voice cuts erasure without hurting meaning.

This matters because "helpful" rewriting tools can standardize identity-bearing language without anyone noticing.

Key points

  • The marker lexicon has 108 items: Indian English 52, Singaporean 32, Nigerian 24. Each item is typed as lexical, pragmatic or syntactic. The corpus holds 624 marker instances, detected by word-boundary-aware pattern matching (Cohen's κ = 0.89 on 500 items).
  • The models were Mistral-7B-Instruct, Apertus-8B-Instruct, DeepHat-7B, MiMo-7B and Qwen3-8B.
    • Under the baseline prompt, IER ranged from 3.5% (Qwen3-8B) to 20.5% (Mistral-7B), a 5.9× spread.
    • Mistral had the highest SPS (0.857) but the worst retention; Qwen3 showed the reverse.
    • Parameter count showed no link to erasure.
  • Overall IER was 10.26% and mean SPS 0.748. The two scores are nearly independent (R² = 0.061).
  • Pragmatic markers (e.g., "kindly", "lah", "Respected sir") were erased 71.5% of the time versus 37.1% for lexical ones (1.9×). A preservation-oriented prompt lowered erasure by 29% with no semantic loss.
  • Two algorithmic fixes were tested:
    • Copy-constrained decoding on tagged marker spans (200 texts) cut IER by 47%, keeping SPS within 2%.
    • Reranking 5 candidates by SPS − 0.3×IER cut IER by 31%.
    • Validation used proxies only: 91% agreement with existing annotations, 89% with an LLM judge, and r = 0.94 against an alternate encoder (M-USE).

Interesting ideas

  • I1. Score a rewrite on two independent axes, meaning kept versus form-bearing markers kept, because a single similarity score hides the loss of how something was said.
  • I2. Treat variety-specific items like "do the needful", "prepone" or "discuss about" as legitimate productive constructions with a typed lexicon entry, not as errors to fix.
  • I3. Pragmatic markers suffer most because their function (face-work, hierarchy, solidarity) cannot be recovered from propositional content alone.
  • I4. Mark protected spans in the input and force the generator to copy them verbatim, a cheap constraint that roughly halved erasure here.
  • I5. Generate several candidates and pick one with a combined objective that rewards meaning and penalizes marker loss, instead of trusting the first output.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (back): The text → graph → text compressor could report a marker-retention score next to its meaning-fidelity check, so round-trips that flatten variety constructions get flagged.
  • I2 (tokens): The en crate could carry a small World English lexicon of multiword lemmas and particles, so the tagger and lemmatizer label them consistently instead of producing odd analyses.
  • I3 (frames): prag could record "kindly", "lah" or "my dear" as explicit politeness or stance features on the speech-act frame, so this relational layer survives into the meaning graph.
  • I4 (back): Generation and translation through the meta-world could carry a "keep surface form" flag on nodes from marked constructions and realize them literally.
  • I5 (learn): When LLM teacher tiers produce paraphrases or silver data, rerank their outputs with a similar fidelity-minus-erasure score before accepting them.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Tsinghua UniversityBeijing · China415 works · 94,028 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksSafety Research · 626 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksEnglish · 2,379 works

Cite

Cite as Satyam Kumar Navneet, Joydeep Chandra and Yong Zhang. 2026. When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models. arXiv:2602.22145.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1145/3772363.3799085 · arXiv 2602.22145

@misc{navneet2026when,
  title = {{When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models}},
  author = {Satyam Kumar Navneet, Joydeep Chandra and Yong Zhang},
  year = {2026},
  doi = {10.1145/3772363.3799085},
  eprint = {2602.22145},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1145/3772363.3799085}
}

References

30 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Dhruv Agarwal, Mor Naaman, and Aditya Vashistha. 2025. AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1117, 21 pages. doi:10.1145/3706598.3713564 ↗
  2. Jeffrey Basoah, Daniel Chechelnitsky, Tao Long, Katharina Reinecke, Chrysoula Zerva, Kaitlyn Zhou, Mark Díaz, and Maarten Sap. 2025. Not Like Us, Hunty: Measuring Perceptions and Behavioral Effects of Minoritized Anthropomorphic Cues in LLMs. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 710–745. doi:10.1145/3715275.3732045 ↗
  3. Gábor Bella, Paula Helm, Gertraud Koch, and Fausto Giunchiglia. 2024. Tackling Language Modelling Bias in Support of Linguistic Diversity. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (Rio de Janeiro, Brazil) (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 562–572. doi:10.1145/3630106.3658925 ↗
  4. Tuhin Chakrabarty, Philippe Laban, and Chien-Sheng Wu. 2025. Can AI writing be salvaged? Mitigating Idiosyncrasies and Improving Human-AI Alignment in the Writing Process through Edits. arXiv:2409.14509 [cs.CL] https://arxiv.org/ abs/2409.14509 ↗
  5. Joydeep Chandra, Aleksandr Algazinov, Satyam Kumar Navneet, Rim El Filali, Matt Laing, Andrew Hanna, and Yong Zhang. 2026. TRACE: Transparent Web Reliability Assessment with Contextual Explanations. arXiv:2506.12072 [cs.IR] https://arxiv.org/abs/2506.12072 ↗
  6. Joydeep Chandra, Ramanjot Kaur, and Rashi Sahay. 2025. Integrated Framework for Equitable Healthcare AI: Bias Mitigation, Community Participation, and Regulatory Governance. In 2025 IEEE 14th International Conference on Communication Systems and Network Technologies (CSNT). 819–825. doi:10.1109/CSNT64827.2025. 10968102 ↗
  7. Joydeep Chandra and Prabal Manhas. 2024. Adversarial Robustness in Optimized LLMs: Defending Against Attacks. SSRN Electronic Journal (01 12 2024). doi:10. 2139/ssrn.5116078 ↗
  8. Joydeep Chandra, Prabal Manhas, Ramanjot Kaur, and Rashi Sahay. 2025. A Unified Approach to Large Language Model Optimization: Methods, Metrics, and Benchmarks (1st ed.). CRC Press. 7 pages. find
  9. Ginel Dorleon and Shirin Shujaa. 2025. Uncovering Cultural Biases and Stereotypes in Large Language Models. In Intelligence and Equity: Shaping the Future of Knowledge: 27th International Conference on Asian Digital Libraries, ICADL 2025, Metro Manila, Philippines, December 3-5, 2025, Proceedings (Metro Manila, Philippines). Springer-Verlag, Berlin, Heidelberg, 64–77. doi:10.1007/978-981-95-4861-3_5 ↗
  10. Rebecca Dorn, Lee Kezar, Fred Morstatter, and Kristina Lerman. 2024. Harmful Speech Detection by Language Models Exhibits Gender-Queer Dialect Bias. In Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (San Luis Potosi, Mexico) (EAAMO ’24). Association for Computing Machinery, New York, NY, USA, Article 6, 12 pages. doi:10.1145/3689904.3694704 ↗
  11. Said Fathi. 2024. Revisiting Brown and Levinson’s Theory of Politeness. European Journal of Language and Culture Studies 3, 5 (Sept. 2024), 1–11. doi:10.24018/ ejlang.2024.3.5.137 ↗
  12. Md Meftahul Ferdaus, Mahdi Abdelguerfi, Elias Loup, Kendall N. Niles, Ken Pathak, and Steven Sloan. 2026. Towards Trustworthy AI: A Review of Ethical and Robust Large Language Models. ACM Comput. Surv. 58, 7, Article 176 (Jan. 2026), 43 pages. doi:10.1145/3777382 ↗
  13. Vinitha Gadiraju, Shaun Kane, Sunipa Dev, Alex Taylor, Ding Wang, Remi Denton, and Robin Brewer. 2023. "I wouldn’t say offensive but...": Disability-Centered Perspectives on Large Language Models. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (Chicago, IL, USA) (FAccT ’23). Association for Computing Machinery, New York, NY, USA, 205–216. doi:10.1145/3593013.3593989 ↗
  14. Rebekka Görge, Michael Mock, and Héctor Allende-Cid. 2025. Detecting Linguistic Indicators for Stereotype Assessment with Large Language Models. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 2796–2814. doi:10.1145/3715275.3732181 ↗
  15. Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pasztor, Bettina Messmer, Dhia Garbaya, Eduard Frank Ďurech, Ido Hakimi, Juan García Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolčec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, Ines Altemir Marinas, Mohammad Hossein Amani, Matin Ansaripour, Ilia Badanin, Harold Benoit, Emanuela Boros, Nicholas Browning, Fabian Bösch, Maximilian Böther, Niklas Canova, Camille Challier, Clement Charmillot, Jonathan Coles, Jan Deriu, Arnout Devos, Lukas Drescher, Daniil Dzenhaliou, Maud Ehrmann, Dongyang Fan, Simin Fan, Silin Gao, Miguel Gila, María Grandury, Diba Hashemi, Alexander Hoyle, Jiaming Jiang, Mark Klein, Andrei Kucharavy, Anastasiia Kucherenko, Frederike Lübeck, Roman Machacek, Theofilos Manitaras, Andreas Marfurt, Kyle Matoba, Simon Matrenok, Henrique Mendoncça, Fawzi Roberto Mohamed, Syrielle Montariol, Luca Mouchel, Sven Najem-Meyer, Jingwei Ni, Gennaro Oliva, Matteo Pagliardini, Elia Palme, Andrei Panferov, Léo Paoletti, Marco Passerini, Ivan Pavlov, Auguste Poiroux, Kaustubh Ponkshe, Nathan Ranchin, Javi Rando, Mathieu Sauser, Jakhongir Saydaliev, Muhammad Ali Sayfiddinov, Marian Schneider, Stefano Schuppli, Marco Scialanga, Andrei Semenov, Kumar Shridhar, Raghav Singhal, Anna Sotnikova, Alexander Sternfeld, Ayush Kumar Tarun, Paul Teiletche, Jannis Vamvas, Xiaozhe Yao, Hao Zhao Alexander Ilic, Ana Klimovic, Andreas Krause, Caglar Gulcehre, David Rosenthal, Elliott Ash, Florian Tramèr, Joost VandeVondele, Livio Veraldi, Martin Rajman, Thomas Schulthess, Torsten Hoefler, Antoine Bosselut, Martin Jaggi, and Imanol Schlag. 2025. Apertus: Democratizing Open and Compliant LLMs for Global Language Environments. https://arxiv.org/abs/2509.14233. ↗
  16. Saurabh Khanna and Xinxu Li. 2025. Invisible Languages of the LLM Universe. arXiv:2510.11557 [cs.CL] https://arxiv.org/abs/2510.11557 ↗
  17. Cass Mayeda, Arinjay Singh, Arnav Mahale, Laila Shereen Sakr, and Mai ElSherief. 2025. Applying Data Feminism Principles to Assess Bias in English and Arabic NLP Research. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 1769–1792. doi:10.1145/3715275.3732119 ↗
  18. Roberto Navigli, Simone Conia, and Björn Ross. 2023. Biases in Large Language Models: Origins, Inventory, and Discussion. J. Data and Information Quality 15, 2, Article 10 (June 2023), 21 pages. doi:10.1145/3597307 ↗
  19. Anna Neumann, Elisabeth Kirsten, Muhammad Bilal Zafar, and Jatinder Singh. 2025. Position is Power: System Prompts as a Mechanism of Bias in Large Language Models (LLMs). In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 573–598. doi:10.1145/3715275.3732038 ↗
  20. Charles Nimo, Irfan Essa, and Michael Best. 2025. Africa Health Check: Probing Cultural Bias in Medical LLMs. In Proceedings of the 5th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ’25). Association for Computing Machinery, New York, NY, USA, 289. doi:10.1145/ 3757887.3767687 find
  21. Nassim Parvin. 2025. Expression and Erasure: AI, English, and the Shaping of Digital Futures. In Adjunct Proceedings of the Sixth Decennial Aarhus Conference: Computing X Crisis (AAR Adjunct ’25). Association for Computing Machinery, New York, NY, USA, Article 12, 4 pages. doi:10.1145/3737609.3747117 ↗
  22. Shirin Shujaa, Ginel Dorleon, and Arthur Tang. 2025. How LLMs Handle Cultural Bias: Reactions to Asian Minority Historical Narratives. In Intelligence and Equity: Shaping the Future of Knowledge: 27th International Conference on Asian Digital Libraries, ICADL 2025, Metro Manila, Philippines, December 3-5, 2025, Proceedings (Metro Manila, Philippines). Springer-Verlag, Berlin, Heidelberg, 39–52. doi:10. 1007/978-981-95-4861-3_3 find
  23. Mozhgan Talebpour, Yunfei Long, Alba G. Seco De Herrera, and Shoaib Jameel. 2025. Bias in Language Models: Interplay of Architecture and Data?. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (Padua, Italy) (SIGIR ’25). Association for Computing Machinery, New York, NY, USA, 2637–2641. doi:10.1145/3726302.3730172 ↗
  24. DeepHat Team. 2024. DeepHat-V1-7B: A Cybersecurity and DevOps Fine-tune of Qwen2.5-Coder. https://huggingface.co/DeepHat/DeepHat-V1-7B. ↗
  25. Mistral AI Team. 2024. Mistral-7B-Instruct-v0.3. https://huggingface.co/mistralai/ Mistral-7B-Instruct-v0.3. ↗
  26. Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https://arxiv.org/abs/2505.09388 ↗
  27. Azmine Toushik Wasi, Raima Islam, Mst Rafia Islam, Farig Sadeque, Taki Hasan Rafi, and Dong-Kyu Chae. 2025. Dialectal Bias in Bengali: An Evaluation of Multilingual Large Language Models Across Cultural Variations. In Companion Proceedings of the ACM on Web Conference 2025 (Sydney NSW, Australia) (WWW ’25). Association for Computing Machinery, New York, NY, USA, 1380–1384. doi:10.1145/3701716.3715468 ↗
  28. Yisong Xiao, Aishan Liu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. 2025. Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models. Proc. ACM Softw. Eng. 2, ISSTA, Article ISSTA012 (June 2025), 24 pages. doi:10.1145/3728881 ↗
  29. LLM-Core-Team Xiaomi. 2025. MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining. arXiv:2505.07608 [cs.CL] https://arxiv.org/abs/2505.07608 ↗
  30. Shiyue Zhang, Asli Celikyilmaz, Jianfeng Gao, and Mohit Bansal. 2021. EmailSum: Abstractive Email Thread Summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/navneet-2026-ai-writes-whose-voice-remains/