---
title: "From a River in Gilead to the Inference Distributions of Large"
description: "Alignment has mostly removed openly racist wording from LLM outputs, but attitudes can persist in the probabilities a model… 36 references."
authors: ["Chowdhury Mohammad Abdullah", "Rita Orji"]
year: 2026
venue: "arXiv:2609.18068"
source: "https://arxiv.org/abs/2609.18068"
doi: "10.48550/arxiv.2609.18068"
arxiv: "2609.18068"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/abdullah-2026-river-gilead-inference-distributions-large/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abdullah-2026-river-gilead-inference-distributions-large/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/30/) / From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale

# From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale

**Authors:** Chowdhury Mohammad Abdullah and Rita Orji · **Year:** 2026 · **Venue:** [arXiv:2609.18068](https://md.research.elamur.ai/free/en/venues/proceedings-of-the-aaai-acm-conference-on-ai-ethics-and-society/) ([Proceedings of the AAAI/ACM Conference on AI Ethics and Society 2026](https://md.research.elamur.ai/free/en/venues/proceedings-of-the-aaai-acm-conference-on-ai-ethics-and-society/2026/)) · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2609.18068) · [PDF](https://arxiv.org/pdf/2609.18068) · [DOI 10.48550/arxiv.2609.18068](https://doi.org/10.48550/arxiv.2609.18068) · [arXiv 2609.18068](https://arxiv.org/abs/2609.18068) · [OpenAlex](https://openalex.org/W7213519679)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Alignment has mostly removed openly racist wording from LLM outputs, but attitudes can persist in the probabilities a model assigns. Audits that read only generated text miss this. The authors extend the computational matched-guise method of Hofmann et al. (2024), which contrasted SAE with AAVE, by adding two Nigerian varieties. Nigerian Standard English (NSE) carries domestic prestige, while Nigerian Pidgin (NP) is stigmatized. Because the two share national and racial identity but differ in prestige, they help separate identity effects from prestige effects. The study uses 260 meaning-equivalent sentence quadruples, one version per variety. Ten open-weight LLMs are scored by log-probability on positive and negative housing adjectives in three settings of increasing intimacy: landlord screening, neighbor acceptance and roommate choice. In every model, AAVE and NP received more negative adjectives than SAE, and NP was penalized most. Each variety triggered its own stereotype cluster. NSE beat SAE for tenancy but lost ground as relationships became closer. The bias therefore runs along both race and prestige, so pre-deployment audits need to be dialect-aware.

## Key points
- The study compares four varieties: SAE, AAVE, NSE and NP. It is the first covert-bias probe for a Black African postcolonial English-speaking population. The paper appeared at AIES 2026.
- The corpus has 260 meaning-matched quadruples. NSE and NP versions were produced by an LLM and then checked by two native speakers of Nigerian English for meaning equivalence and authenticity.
- Bias is measured as q-scores, which give each item's deviation from the SAE baseline for the same sentence and adjective. Ten open-weight models were tested, covering varied origins and both base and instruction-tuned variants.
- AAVE and NP were penalized in all ten models and all three housing contexts, with NP penalized most. NSE was favored in tenant screening but penalized as social closeness increased.
- The authors acknowledge several limits. Probability bias may correlate only weakly with downstream harm, base and instruction-tuned models were not compared, and the adjective lexicons were built without input from Nigerian communities. Data and code are released on GitHub (intricate-potato/dialect-prejudice-of-llms).

## Interesting ideas
- **Meaning-matched variants across dialects.** Holding content fixed and changing only the variety means any change in model behaviour can be attributed to dialect alone. This is a clean invariance test for any NLP component.
- **Scoring against a per-item baseline.** Measuring each item as a deviation from its own SAE counterpart makes the aggregate result robust to noise in the translations. Only systematic distortion can bias the result.
- **A third condition to break a confound.** NSE was added because SAE and AAVE differ in race and prestige at once, and NSE lets the two factors be pulled apart. Adding a condition that varies one factor alone is a reusable experimental design.
- **Auditing internal probabilities instead of surface output.** A system can produce polite text while its internal scores still encode a skew, so the scores themselves should be inspected.
- **A graded context ladder.** Testing the same judgment in tenant, neighbor and roommate settings showed that bias can reverse as context changes.

## For Mova

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

- **I1** (tree): Parse AAVE- or NP-style renderings of the same sentence and check that the UD trees collapse to the same core relations after dialect conversion; mismatches show where the parser or the conversion fails.
- **I2** (learn): Evaluate repair rules or teacher-tier silver data as paired deltas against an untouched control version of the same sentence, rather than as raw accuracy.
- **I3** (learn): When an error correlates with two features, such as register and genre, build negative-control sets that vary only one of them before blaming either.
- **I4** (answer): Log the confidence each reader assigns to competing answers across dialectal paraphrases of a test question, so hidden preferences show up even when the final answer is correct.
- **I5** (frames): Test implicature or irony detection on the same utterance placed in progressively more personal or formal frames, to check whether speech-act labels drift with social setting rather than with content.

*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

- [Dalhousie University](https://md.research.elamur.ai/free/en/universities/dalhousie-university/) — University, Halifax, Canada · 9 works · 480 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Safety Research](https://md.research.elamur.ai/free/en/shelves/social-sciences/safety-research/) · 543 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works

## Cite

Cite as: Chowdhury Mohammad Abdullah and Rita Orji. 2026. From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale. arXiv:2609.18068.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{abdullah2026river,
  title = {{From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale}},
  author = {Chowdhury Mohammad Abdullah and Rita Orji},
  year = {2026},
  doi = {10.48550/arxiv.2609.18068},
  eprint = {2609.18068},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.18068}
}
```

- [← previous Frequency modulates structural choice in Turkish suspended affixation: a latent-process account](https://md.research.elamur.ai/free/en/papers/turk-2026-frequency-modulates-structural-choice-turkish/)
- [next → From Annotated Clinical Narratives to Ontology: Structuring Brazilian Portuguese Clinical Data](https://md.research.elamur.ai/free/en/papers/oliveira-2026-annotated-clinical-narratives-ontology-structuring/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/30/)

## References

36 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Akaabre, P. B.; Hackworth, J.; and Keckesova, N. 2025. Anti-black rental housing discrimination in the multicultural city? A field experiment in Toronto, Canada. Journal of Urban Affairs, 1–15. [find](https://scholar.google.com/scholar?q=Akaabre%2C+P%2E+B%2E%3B+Hackworth%2C+J%2E%3B+and+Keckesova%2C+N%2E+2025%2E+Anti%2Dblack+rental+housing+discrimination+in+the+multicultural+city%3F+A+field+experiment+in+Toronto%2C+Canada%2E+Journal+of+Urban+Affairs%2C+1%E2%80%9315%2E)
2. Alvero, A.; Lee, J.; Regla-Vargas, A.; Kizilcec, R. F.; Joachims, T.; and Antonio, A. L. 2024. Large language models, social demography, and hegemony: comparing authorship in human and synthetic text. Journal of Big Data, 11(1): 138. [↗](https://doi.org/10.1186/s40537-024-00986-7)
3. Baugh, J. 2021. Linguistic Profiling and Language-Based Discrimination. Oxford Bibliographies. [↗](https://doi.org/10.1093/obo/9780199772810-0267)
4. Bender, E. M.; Gebru, T.; McMillan-Major, A.; and Shmitchell, S. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’21, 610–623. New York, NY, USA: Association for Computing Machinery. ISBN 9781450383097. [find](https://scholar.google.com/scholar?q=Bender%2C+E%2E+M%2E%3B+Gebru%2C+T%2E%3B+McMillan%2DMajor%2C+A%2E%3B+and+Shmitchell%2C+S%2E+2021%2E+On+the+Dangers+of+Stochastic+Parrots%3A+Can+Language+Models+Be+Too+Big%3F+In+Proceedings+of+the+2021+ACM+Conference+on+Fairness%2C+Acco)
5. Blodgett, S. L.; Barocas, S.; Daumé III, H.; and Wallach, H. 2020. [Language (Technology) is Power: A Critical Survey of “Bias” in NLP](https://md.research.elamur.ai/free/en/papers/blodgett-2020-language-technology-is-power-critical/). In Jurafsky, D.; Chai, J.; Schluter, N.; and Tetreault, J., eds., Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5454–5476. Online: Association for Computational Linguistics.
6. Fiske, S. T.; Cuddy, A. J.; Glick, P.; and Xu, J. 2018. A model of (often mixed) stereotype content: Competence and warmth respectively follow from perceived status and competition. In Social cognition, 162–214. Routledge. [find](https://scholar.google.com/scholar?q=Fiske%2C+S%2E+T%2E%3B+Cuddy%2C+A%2E+J%2E%3B+Glick%2C+P%2E%3B+and+Xu%2C+J%2E+2018%2E+A+model+of+%28often+mixed%29+stereotype+content%3A+Competence+and+warmth+respectively+follow+from+perceived+status+and+competition%2E+In+Social+cognitio)
7. Fleisig, E.; Smith, G.; Bossi, M.; Rustagi, I.; Yin, X.; and Klein, D. 2024. [Linguistic Bias in ChatGPT: Language Models Reinforce Dialect Discrimination](https://md.research.elamur.ai/free/en/papers/fleisig-2024-linguistic-bias-chatgpt-language-models/). In Al-Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 13541–13564. Miami, Florida, USA: Association for Computational Linguistics.
8. Gaddis, S. M.; and Ghoshal, R. 2020. Searching for a Roommate: A Correspondence Audit Examining Racial/Ethnic and Immigrant Discrimination among Millennials. Socius, 6: 2378023120972287. PMID: 34355061. [↗](https://doi.org/10.1177/2378023120972287)
9. Gallegos, I. O.; Rossi, R. A.; Barrow, J.; Tanjim, M. M.; Kim, S.; Dernoncourt, F.; Yu, T.; Zhang, R.; and Ahmed, N. K. 2024. [Bias and Fairness in Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/gallegos-2023-bias-fairness-large-language-models/). Computational Linguistics, 50(3): 1097–1179. [↗](https://doi.org/10.1162/coli_a_00524)
10. Galster, G. 1990. Racial discrimination in housing markets during the 1980s: a review of the audit evidence. Journal of Planning Education and Research, 9(3): 165–175. [↗](https://doi.org/10.1177/0739456x9000900302)
11. Greer, T. W.; Mills, M. T.; and Lapka, S. 2024. The influence of linguistic profiling on perceived employability: a new application of the systems theory framework of career development. Career development international, 29(3): 352–366. [↗](https://doi.org/10.1108/cdi-07-2023-0231)
12. Groenwold, S.; Ou, L.; Parekh, A.; Honnavalli, S.; Levy, S.; Mirza, D.; and Wang, W. Y. 2020. [Investigating African-American Vernacular English in Transformer-Based Text Generation](https://md.research.elamur.ai/free/en/papers/groenwold-2020-investigating-african-american-vernacular-english/). In Webber, B.; Cohn, T.; He, Y.; and Liu, Y., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 5877– 5883. Online: Association for Computational Linguistics.
13. Halevy, M.; Harris, C.; Bruckman, A.; Yang, D.; and Howard, A. 2021. [Mitigating Racial Biases in Toxic Language Detection with an Equity-Based Ensemble Framework](https://md.research.elamur.ai/free/en/papers/halevy-2021-mitigating-racial-biases-toxic-language/). In Proceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, EAAMO ’21. New York, NY, USA: Association for Computing Machinery. ISBN 9781450385534.
14. Himelstein, R.; LeVi, A.; Youngmann, B.; Nemcovsky, Y.; and Mendelson, A. 2026. Silenced biases: The dark side LLMs learned to refuse. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 37452–37461. [find](https://scholar.google.com/scholar?q=Himelstein%2C+R%2E%3B+LeVi%2C+A%2E%3B+Youngmann%2C+B%2E%3B+Nemcovsky%2C+Y%2E%3B+and+Mendelson%2C+A%2E+2026%2E+Silenced+biases%3A+The+dark+side+LLMs+learned+to+refuse%2E+In+Proceedings+of+the+AAAI+Conference+on+Artificial+Intelligence%2C)
15. Hofmann, V.; Kalluri, P. R.; Jurafsky, D.; and King, S. 2024. AI generates covertly racist decisions about people based on their dialect. Nature, 633(8028): 147–154. [↗](https://doi.org/10.1038/s41586-024-07856-5)
16. Hosseini, S.; and Seilani, H. 2025. The role of agentic AI in shaping a smart future: A systematic review. Array, 26: 100399. [↗](https://doi.org/10.1016/j.array.2025.100399)
17. Jemiluyi, O. 2025. From Naija to the States: stereotypes and resistance of the Nigerian man in America. Howard Journal of Communications, 1–22. [↗](https://doi.org/10.1080/10646175.2025.2525258)
18. Jones, T.; Kalbfeld, J. R.; Hancock, R.; and Clark, R. 2019. Testifying while black: An experimental study of court reporter accuracy in transcription of African American English. Language, 95(2): e216–e252. [↗](https://doi.org/10.1353/lan.2019.0042)
19. Kachru, B. B. 1990. World Englishes and applied linguistics. World Englishes, 9(1): 3–20. [↗](https://doi.org/10.1111/j.1467-971x.1990.tb00683.x)
20. Lambert, W. E.; Hodgson, R. C.; Gardner, R. C.; and Fillenbaum, S. 1960. Evaluational reactions to spoken languages. The journal of abnormal and social psychology, 60(1): 44. [↗](https://doi.org/10.1037/h0044430)
21. Lev-Ari, S.; and Keysar, B. 2010. Why don’t we believe non-native speakers? The influence of accent on credibility. Journal of Experimental Social Psychology, 46(6): 1093– 1096. [find](https://scholar.google.com/scholar?q=Lev%2DAri%2C+S%2E%3B+and+Keysar%2C+B%2E+2010%2E+Why+don%E2%80%99t+we+believe+non%2Dnative+speakers%3F+The+influence+of+accent+on+credibility%2E+Journal+of+Experimental+Social+Psychology%2C+46%286%29%3A+1093%E2%80%93+1096%2E)
22. Lin, Z.; Guan, S.; Zhang, W.; Zhang, H.; Li, Y.; and Zhang, H. 2024. Towards trustworthy LLMs: a review on debiasing and dehallucinating in large language models. Artificial Intelligence Review, 57(9): 243. [↗](https://doi.org/10.1007/s10462-024-10896-y)
23. Lippi-Green, R. 1997. English with an accent: Language, ideology, and discrimination in the United States. Psychology Press. [find](https://scholar.google.com/scholar?q=Lippi%2DGreen%2C+R%2E+1997%2E+English+with+an+accent%3A+Language%2C+ideology%2C+and+discrimination+in+the+United+States%2E+Psychology+Press%2E)
24. Massey, D. S.; and Lundy, G. 2001. Use of Black English and racial discrimination in urban housing markets: New methods and findings. Urban affairs review, 36(4): 452– 469. [↗](https://doi.org/10.1177/10780870122184957)
25. McClintock, E. A. 2010. When does race matter? Race, sex, and dating at an elite university. Journal of Marriage and Family, 72(1): 45–72. [find](https://scholar.google.com/scholar?q=McClintock%2C+E%2E+A%2E+2010%2E+When+does+race+matter%3F+Race%2C+sex%2C+and+dating+at+an+elite+university%2E+Journal+of+Marriage+and+Family%2C+72%281%29%3A+45%E2%80%9372%2E)
26. Pallavi, L.; Rishika, D.; Sree, G. B. S.; Ruthwik, J.; and Yashaswini, J. 2025. SmartRent-AI Driven Online House Rental Platform. In 2025 5th International Conference on Soft Computing for Security Applications (ICSCSA), 346– 353. IEEE. [find](https://scholar.google.com/scholar?q=Pallavi%2C+L%2E%3B+Rishika%2C+D%2E%3B+Sree%2C+G%2E+B%2E+S%2E%3B+Ruthwik%2C+J%2E%3B+and+Yashaswini%2C+J%2E+2025%2E+SmartRent%2DAI+Driven+Online+House+Rental+Platform%2E+In+2025+5th+International+Conference+on+Soft+Computing+for+Security+Ap)
27. Purnell, T.; Idsardi, W.; and Baugh, J. 1999. Perceptual and phonetic experiments on American English dialect identification. Journal of language and social psychology, 18(1): 10–30. [↗](https://doi.org/10.1177/0261927x99018001002)
28. Salazar, J.; Liang, D.; Nguyen, T. Q.; and Kirchhoff, K. 2020. [Masked Language Model Scoring](https://md.research.elamur.ai/free/en/papers/salazar-2020-masked-language-model-scoring/). In Jurafsky, D.; Chai, J.; Schluter, N.; and Tetreault, J., eds., Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2699–2712. Online: Association for Computational Linguistics.
29. Smiley, C.; and Fakunle, D. 2018. From “brute” to “thug:” The demonization and criminalization of unarmed Black male victims in America. In Police and the unarmed Black male crisis, 8–39. Routledge. [find](https://scholar.google.com/scholar?q=Smiley%2C+C%2E%3B+and+Fakunle%2C+D%2E+2018%2E+From+%E2%80%9Cbrute%E2%80%9D+to+%E2%80%9Cthug%3A%E2%80%9D+The+demonization+and+criminalization+of+unarmed+Black+male+victims+in+America%2E+In+Police+and+the+unarmed+Black+male+crisis%2C+8%E2%80%9339%2E+Routledge%2E)
30. Squires, G. D.; and Chadwick, J. 2006. Linguistic profiling: A continuing tradition of discrimination in the home insurance industry? Urban Affairs Review, 41(3): 400–415. [find](https://scholar.google.com/scholar?q=Squires%2C+G%2E+D%2E%3B+and+Chadwick%2C+J%2E+2006%2E+Linguistic+profiling%3A+A+continuing+tradition+of+discrimination+in+the+home+insurance+industry%3F+Urban+Affairs+Review%2C+41%283%29%3A+400%E2%80%93415%2E)
31. Tanyi, J. N.; Ekoro, D. E.; Okpene, D. E.; and Edung, A. 2025. The Sociolinguistics of Nigerian Pidgin (Naija) in the Digital Age: A Study of Its Evolving Status, Standardization, and Functions on Social Media Platforms. Performance: Journal of Law and Humanities, 3(1): 60–68. [find](https://scholar.google.com/scholar?q=Tanyi%2C+J%2E+N%2E%3B+Ekoro%2C+D%2E+E%2E%3B+Okpene%2C+D%2E+E%2E%3B+and+Edung%2C+A%2E+2025%2E+The+Sociolinguistics+of+Nigerian+Pidgin+%28Naija%29+in+the+Digital+Age%3A+A+Study+of+Its+Evolving+Status%2C+Standardization%2C+and+Functions+on+Soc)
32. Tatman, R. 2017. [Gender and Dialect Bias in YouTube’s Automatic Captions](https://md.research.elamur.ai/free/en/papers/tatman-2017-gender-dialect-bias-youtube-s/). In Hovy, D.; Spruit, S.; Mitchell, M.; Bender, E. M.; Strube, M.; and Wallach, H., eds., Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, 53–59. Valencia, Spain: Association for Computational Linguistics.
33. Tucker, G. R.; and Lambert, W. E. 1969. White and Negro Listeners’ Reactions to Various American-English Dialects*. Social Forces, 47(4): 463–468. [↗](https://doi.org/10.2307/2574535)
34. Ugwuanyi, K. O.; and Aboh, S. C. 2026. Nigerian English: History, functions and features. World Englishes, 45(1): 7– 18. [↗](https://doi.org/10.1111/weng.12732)
35. Wen, X.; Li, B.; Huang, T.; and Chen, M. 2024. Red Teaming Language Models for Processing Contradictory Dialogues. In Al-Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 11611–11630. Miami, Florida, USA: Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Wen%2C+X%2E%3B+Li%2C+B%2E%3B+Huang%2C+T%2E%3B+and+Chen%2C+M%2E+2024%2E+Red+Teaming+Language+Models+for+Processing+Contradictory+Dialogues%2E+In+Al%2DOnaizan%2C+Y%2E%3B+Bansal%2C+M%2E%3B+and+Chen%2C+Y%2E%2DN%2E%2C+eds%2E%2C+Proceedings+of+the+2024+Confere)
36. Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Funtowicz, M.; Davison, J.; Shleifer, S.; von Platen, P.; Ma, C.; Jernite, Y.; Plu, J.; Xu, C.; Le Scao, T.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. 2020. [Transformers: State-of-the-Art Natural Language Processing](https://md.research.elamur.ai/free/en/papers/wolf-2020-transformers-state-art-natural-language/). In Liu, Q.; and Schlangen, D., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 38–45. Online: Association for Computational Linguistics.

*Updated 1 November 2026*