Elamur Research⌕
☰
home / papers / 2026 / AI systems and the reproduction of (standard) language ideologies in World Englishes

AI systems and the reproduction of (standard) language ideologies in World Englishes

Paper · 2026 · Kingsley Ugwuanyi

Summary

This position paper asks how AI systems carry forward the belief that one standard, Inner Circle English is the correct form and other Englishes are deviant. It uses a sociolinguistic lens to examine prior empirical studies, media coverage, social media debates and sample model outputs. The author traces this ideology through several layers of the AI pipeline: web-scale training corpora, RLHF annotation guidelines, benchmark design, user feedback and public commentary. One case study is the online backlash against the word delve as a supposed marker of AI-written text. The paper argues that this backlash lets Global North speakers police Global South usage, even though corpus evidence does not support the attribution. Following Christian Mair, it describes a "standardisation paradox": LLMs push English toward uniformity while also absorbing variety from diverse data and from annotators in the Global South. The paper matters because these biases have material effects on hiring, education and transcription. It calls for parallel multi-variety datasets, a review of data-cleaning steps, and dialect-aware model components.

Key points

  • Fleisig et al. (2024) tested GPT-4 and GPT-3.5 on 10 English varieties. Relative to Standard American English, prompts in other varieties produced a 19% increase in stereotyping, a 25% rise in demeaning responses and 9% more cases where the model failed to understand.
  • Lin et al. (2024) gave reasoning tasks in AAVE to GPT-4o/4/3.5-turbo, LLaMA-3.1/3, Mistral and Phi-3. Even when the AAVE prompts were equivalent to Standard American English versions, nearly all models became brittle and unfair.
  • Raji et al. (2021) are cited as showing that GLUE and SuperGLUE reflect a narrow slice of American English.
  • ImageNet tried to be multilingual, yet over 45% of its images come from the US. China and India together contribute roughly 3%.
  • Asked for an abstract in Nigerian English, ChatGPT-4 produced a caricature leaning on Nigerian Pidgin. When asked for British English, it tends to drift back to American norms.
  • Held et al. (2023) Multi-VALUE uses swappable adapters to tune existing models to particular dialects without retraining them from scratch.

Interesting ideas

  • I1. Evaluating a system on meaning-equivalent inputs written in different varieties isolates the dialect gap from task difficulty, which makes bias measurable rather than anecdotal.
  • I2. Small swappable adapters per dialect can widen coverage without retraining the base model, so inclusivity need not cost efficiency.
  • I3. Data cleaning and quality control are a hidden point where variation gets erased, because filters encode one variety's norms as "quality."
  • I4. Output in a requested variety should be checked both for drift toward the default standard and for exaggerated, stereotyped features.
  • I5. The delve case shows that a word's perceived social marking can contradict its actual corpus distribution, so labels should be grounded in counts rather than impressions.

For Mova

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

  • I1 (answer): Build paired standard and non-standard versions of school tests and fable QA items, then report the accuracy gap as its own metric.
  • I2 (tree): gparse could carry per-variety adapter heads, for example for Nigerian or Indian English, on top of one shared parser, selected or blended at parse time.
  • I3 (learn): When witness parsers vote on silver data, log which sentences are discarded and by which variety, and treat a skew as a negative-control warning.
  • I4 (back): Generation tests should check that a British or Nigerian English target keeps its spelling and grammar without padding the text with caricatured markers.
  • I5 (tokens): Any variety or register flags in the en lexicon should be derived from corpus frequencies across varieties, not hand-assigned from intuition.

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

🎓SOAS University of LondonLondon · United Kingdom3 works · 945 citations

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Arts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Kingsley Ugwuanyi. 2026. AI systems and the reproduction of (standard) language ideologies in World Englishes. arXiv:2607.28528.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2607.28528 · arXiv 2607.28528

@misc{ugwuanyi2026ai,
  title = {{AI systems and the reproduction of (standard) language ideologies in World Englishes}},
  author = {Kingsley Ugwuanyi},
  year = {2026},
  doi = {10.48550/arxiv.2607.28528},
  eprint = {2607.28528},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2607.28528}
}

References

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

  1. Aggarwal, Rohit. 2024. Let us "delve" into the influence of Nigerian English on ChatGPT and Global Writing Patterns [LinkedIn post]. 16 April 2024. https://www.linkedin.com/posts/rohitaggarwal_let-us-delve-into-the-influence-of-nigerian-share-7186469322360254464-Grul/ ↗
  2. AI Phrase Finder. 2024. 10 most common ChatGPT words. AI Phrase Finder, 9 March 2024. https://aiphrasefinder.com/common-chatgpt-words/. ↗
  3. Bodunde, Deborah. 2024. Nigerians berate US author Paul Graham for claiming ‘delve’ is only used by ChatGPT. TheCable Lifestyle. 9 April 2024. https://lifestyle.thecable.ng/nigerians-berate-us-author-paul-graham-for-claiming-delve-is-only-used-by-chatgpt/ ↗
  4. Constantino, Tor. 2024. New list ranks AI’s 50 most overused words – It updates monthly. Forbes, 7 October 2024. https://www.forbes.com/sites/torconstantino/2024/10/07/new-list-ranks-ais-50-most-overused-words---updates-monthly/. ↗
  5. Dai, David Wei & Zhu Hua. 2025. When AI meets intercultural communication: New frontiers, new agendas. Applied Linguistics Review 16(2). 747–751. ↗
  6. Davies, Mark. 2013. The News on the Web (NOW) corpus. https://www.english-corpora.org/now/ (accessed 26 February 2026). ↗
  7. Erdocia, Iker, Bettina Migge & Britta Schneider. 2024. Language is not a data set – Why overcoming ideologies of dataism is more important than ever in the age of AI. Journal of Sociolinguistics 28(5). 20–25. ↗
  8. Erdocia, Iker, Britta Schneider & Bettina Migge. 2026. Language in the age of AI technology: From human to non-human authenticity, from public governance to privatised assemblages. Language in Society 55(3). 523–543. ↗
  9. Erfani, Farhang. 2026. Not “just a language model”: AI’s ideological functions. AI & Society 41. 869–881. find
  10. Fleisig, Eve, Genevieve Smith, Madeline Bossi, Ishita Rustagi, Xavier Yin & Dan Klein. 2024. Linguistic bias in ChatGPT: Language models reinforce dialect discrimination.
  11. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 13541–13564. Miami, FL: Association for Computational Linguistics. https://doi.org/10.48550/arXiv.2406.08818. ↗
  12. Graham, Paul [@paulg]. 2024. Someone sent me a cold email proposing a novel project. Then I noticed it used the word “delve.” [Tweet]. Twitter, 7 April 2024. https://x.com/paulg/status/1777030573220933716. ↗
  13. Held, Will, Caleb Ziems & Diyi Yang. 2023. TADA: Task-agnostic dialect adapters for English. In Anna Rogers, Jordan Boyd-Graber & Naoaki Okazaki, Findings of the Association for Computational Linguistics: ACL 2023, 813–824. Toronto: Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-acl.51. ↗
  14. Hern, Alex. 2024. TechScape: How cheap, outsourced labour in Africa is shaping AI English. The Guardian, 16 April 2024. https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt. ↗
  15. Hofmann, Valentin, Pratyusha Ria Kalluri, Dan Jurafsky & Sharese King. 2024. AI generates covertly racist decisions about people based on their dialect. Nature 633(8028). 147– 154. ↗
  16. Irvine, Judith & Susan Gal. 2000. Language ideology and linguistic differentiation. In Paul Kroskrity (ed.), Regimes of language: Ideologies, polities, and identities, 35–83. Santa Fe: School of American Research Press. find
  17. Jones, Rodney. 2025. Culture machines. Applied Linguistics Review 16(2). 753–762. find
  18. Kaubrė, Vytenis. 2024. LLM training data: The 8 main public data sources. Oxylabs blog, 27 September 2024. https://oxylabs.io/blog/llm-training-data. ↗
  19. Lin, Fangru, Shaoguang Mao, Emanuele La Malfa, Valentin Hofmann, Adrian de Wynter, Jing Yao, Si-Qing Chen, Michael Wooldridge & Furu Wei. 2024. One language, many gaps: Evaluating dialect fairness and robustness of large language models in reasoning tasks. arXiv. https://arxiv.org/html/2410.11005v1. ↗
  20. Lippi-Green, Rosina. 2012. English with an accent: Language, ideology and discrimination in the United States, 2nd edn. London: Routledge. find
  21. Mair, Christian. 2023. The study of World Englishes: Impulses from beyond linguistics. Arbeiten aus Anglistik und Amerikanistik / Agenda: Advancing Anglophone Studies 48(1). 15–35. ↗
  22. Mair, Christian. 2025. The digital diffusion of pidgin/creoles: Langwij, kolcha, media and large language models. Journal of Pidgin and Creole Languages. https://doi.org/10.1075/jpcl.24025.mai (advance online publication). ↗
  23. Michel, Shira, Sufi Kaur, Sarah Elizabeth Gillespie, Jeffrey Gleason, Christo Wilson & Avijit Ghosh. 2025. “It’s not a representation of me”: Examining accent bias and digital exclusion in synthetic AI voice services. arXiv. https://arxiv.org/html/2504.09346v1. ↗
  24. Milroy, James & Lesley Milroy. 2012. Authority in language: Investigating standard English. London: Routledge. find
  25. Raji, Inioluwa Deborah, Emily Bender, Amandalynne Paullada, Emily Denton & Alex Hanna. 2021. AI and the everything in the whole wide world benchmark. arXiv. https://doi.org/10.48550/arXiv.2111.15366. ↗
  26. Rodriguez Louro, Celeste. 2025. AI systems are built on English – but not the kind most of the world speaks. The Conversation, 5 May 2025. https://doi.org/10.64628/AA.4mhwhkhr5. ↗
  27. Shankar, Shreya, Yoni Halpern, Eric Breck, James Atwood, Jimbo Wilson & D. Sculley. 2017. No classification without representation: Assessing geodiversity issues in open data sets for the developing world. arXiv. https://doi.org/10.48550/arXiv.1711.08536. ↗
  28. Silverstein, Michael. 1979. Language structure and linguistic ideology. In Paul Clyne, William Hanks & Carol Hofbauer (eds.), The elements: A parasession on linguistic units and levels, 193–247. Chicago: Chicago Linguistic Society. find
  29. Ugwuanyi, Kingsley, Christian Mair, Sender Dovchin, Iker Erdocia, Maria Kuteeva & Esther Airemionkhale. 2026. Generative AI and linguistic diversity in scholarly publishing: Perspectives from World Englishes scholars. Journal of English for Research Publication Purposes 6(2). 266–290. find
  30. Wang, Ding, Shantanu Prabhat & Nithya Sambasivan. 2022. Whose AI dream? In search of the aspiration in data annotation. arXiv. https://doi.org/10.48550/arXiv.2203.10748. ↗
  31. Woolard, Kathryn A. 1998. Introduction: Language Ideology as a field of inquiry. In Schieffelin, Bambi, Kathryn Woolard and Paul Kroskrity (eds.), Language Ideologies: Practice and Theory, 3-47. Oxford: Oxford University Press. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ugwuanyi-2026-ai-systems-reproduction-standard-language/