Elamur Research⌕
☰
home / papers / 2026 / Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek

Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek

Paper · 2026 · Annabella Sakunkoo and Jonathan Sakunkoo

Summary

Most audits of social bias in LLMs use English prompts and romanized names, so it is unclear whether the ethnic rankings they find hold in other languages and writing systems. This student workshop paper tests that assumption by comparing a US-built model (GPT-4o-mini) with a China-built one (DeepSeek-V3). The authors created 45,000 name variants covering 9 ethnic groups and collected 900,000 model judgments of competence and warmth. They varied the prompt language (English, Chinese, Thai) and whether names appeared in Latin or native script. To interpret the results, they introduce a stereotype perceptual map, which places each group in one of four quadrants of the warmth–competence model of stereotypes: admiration, envy, paternalism or contempt.

The two models behave very differently. DeepSeek's math-competence ordering barely changes across settings and always favors Chinese names. GPT reorders groups depending on the script, an effect the authors call script-gated bias. The practical lesson is that fairness checks limited to one language and one script can miss, or even reverse, the hierarchies a model shows in other deployments.

Key points

  • Design:
    • Models: GPT-4o-mini and DeepSeek-V3.
    • Groups: 9 ethnicities (Chinese, Indian, Iranian, Jewish, Russian, Thai, White, Black, Hispanic).
    • Language–script pairs: English with Latin-script names, Chinese with Latin or Chinese script, Thai with Latin or Thai script.
    • Domains: competence and warmth.
    • Scale: 900,000 responses over 45,000 name variations.
  • DeepSeek: It is highly stable in math judgments. Chinese names rank 1st in every condition and Russian names come next. White, Hispanic and Black names sit at the bottom. Black names receive competence deviations beyond −7 points, the largest single-group penalty in the dataset.
  • GPT: Latin-script and native-script conditions form two separate clusters with weak cross-cluster correlation. Transliterating names can sometimes produce rankings inversely correlated with the Latin-script results.
  • White names in GPT: These are never substantially penalized in any condition (a "floor effect"). This signals a Western-centric default that persists even in non-English prompts.
  • Warmth vs. competence: Warmth rankings are less stable than competence rankings in both models. Some groups move from Admiration to Contempt purely through changes in language or script. Others, such as Iranian names, stay in unfavorable quadrants almost everywhere.

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

🎓Stanford UniversityStanford · United States434 works · 314,419 citations🎓University of OxfordOxford · United Kingdom104 works · 157,183 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSafety Research · 543 worksSociology and Political Science · 577 works

🧭 Topics 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksMultilingual & cross-lingual NLP · 828 worksBias, fairness & safety · 517 works

Cite

Cite as Annabella Sakunkoo and Jonathan Sakunkoo. 2026. Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.96

@inproceedings{sakunkoo2026through,
  title = {{Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek}},
  author = {Annabella Sakunkoo and Jonathan Sakunkoo},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.96},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.96}
}

References

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

  1. Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES ’21, page 298–306, New York, NY, USA. Association for Computing Machinery.
  2. Gordon W. Allport. 1954. The Nature of Prejudice. "Addison-Wesley". find
  3. Haozhe An, Christabel Acquaye, Colin Wang, Zongxia Li, and Rachel Rudinger. 2024. Do large language models discriminate in hiring decisions on the basis of race, ethnicity, and gender? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)… find
  4. Marianne Bertrand and Sendhil Mullainathan. 2004. Are emily and greg more employable than lakisha and jamal? a field experiment on labor market discrimination. American Economic Review, 94(4):991–1013. find
  5. Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183–186. find
  6. Jennifer L. Eberhardt. 2019. Biased: Uncovering the Hidden Prejudice That Shapes What We See, Think, and Do. Viking. find
  7. Susan T Fiske. 2026. Prejudice, discrimination, and stereotyping. http://noba.to/jfkx7nrd. Accessed: 2026-2-18. ↗
  8. Susan T Fiske, Amy J C Cuddy, and Peter Glick. 2007. Universal dimensions of social cognition: warmth and competence. Trends Cogn. Sci., 11(2):77–83. find
  9. Susan T Fiske, Amy J C Cuddy, Peter Glick, and Jun Xu. 2002. A model of (often mixed) stereotype content: competence and warmth respectively follow from perceived status and competition. J. Pers. Soc. Psychol., 82(6):878–902. find
  10. Anthony G. Greenwald, Debbie E. McGhee, and Jordan L. K. Schwartz. 1998. Measuring individual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74(6):1464–1480. find
  11. Anna Luca Heimann and Annika Schmitz-Wilhelmy. 2024. Observing interviewees’ inner self: How authenticity cues in job interviews relate to interview and job performance. J. Bus. Psychol. find
  12. Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky, and Sharese King. 2024. AI generates covertly racist decisions about people based on their dialect. Nature, 633(8028):147–154. find
  13. Sullam Jeoung, Jana Diesner, and Halil Kilicoglu. 2023. Examining the causal impact of first names on language models: The case of social commonsense reasoning. In Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023), pages 61–72, Toronto, Canada. Association fo… find
  14. D Katz and K Braly. 1933. Racial stereotypes of one hundred college students. The Journal of Abnormal and Social Psychology, 28(3):280–290. find
  15. F W Kerche, M Zook, and M Graham. 2026. The silicon gaze: A typology of biases and inequality in LLMs through the lens of place. Platforms & Society, 3. find
  16. Mary E Kite, Bernard E Whitley, Jr, and Lisa S Wagner. 2022. Psychology of prejudice and discrimination, 4 edition. Routledge. find
  17. Rowan Hall Maudslay, Hila Gonen, Ryan Cotterell, and Simone Teufel. 2019. It‘s all in the name: Mitigating gender bias with name-based counterfactual data substitution. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conferenc…
  18. Kieran O’Connor, Glenn R Carroll, and Balázs Kovács. 2017. Disambiguating authenticity: Interpretations of value and appeal. PLoS One, 12(6):e0179187. find
  19. Annabella Sakunkoo and Jonathan Sakunkoo. 2025. Name of thrones: How do LLMs rank student names in status hierarchies based on race and gender? In Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025), pages 697–707, Vienna, Austria. Association f… find
  20. Vered Shwartz, Rachel Rudinger, and Oyvind Tafjord. 2020. “you are grounded!”: Latent name artifacts in pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6850–6861, Online. Association for Computational Linguistics. find
  21. Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. 2023. Are emily and greg still more employable than lakisha and jamal? investigating algorithmic hiring bias in the era of chatgpt. CoRR, abs/2310.05135. — [arXiv:2310.05135](https://arxiv.org/abs/2310.05135) find
  22. Jun Wang, Benjamin Rubinstein, and Trevor Cohn. 2022. Measuring and mitigating name biases in neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). find
  23. Robert Wolfe and Aylin Caliskan. 2021. Low frequency names exhibit bias and overfitting in contextualizing language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 518–532, Online and Punta Cana, Dominican Republic. Association for Computation… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sakunkoo-2026-through-looking-glass-multilingual-ai/