Elamur Research⌕
☰
home / papers / 2026 / Side-by-side Comparison Amplifies Dialect Bias in Language Models

Side-by-side Comparison Amplifies Dialect Bias in Language Models

Paper · 2026 · Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel et al.

Summary

The paper asks whether language models judge writers more harshly when their text is in African-American Vernacular English (AAVE) than when it is in Standard American English (SAE), even without any dialect label (covert dialect bias). It also asks whether this changes when the two versions are judged together. The authors use a dataset of paired tweets that carry the same intent in both dialects. Models rate each tweet's content on a 1–5 Likert scale for 12 traits taken from stereotype research. There are two prompting modes. In absolute mode each tweet is rated alone. In contrastive mode the SAE and AAVE versions are rated side by side. An overt variant names the dialect in the prompt. AAVE receives more negative ratings in every condition. The gap grows clearly in contrastive mode and grows further when the dialect is named. Counterfactual fairness finetuning shrinks the gaps for some traits in isolated rating, but the gains are inconsistent in side-by-side rating. This matters because hiring, moderation and similar ranking tasks are comparative by nature, so isolated tests may underestimate the real harm.

Key points

  • The 12 traits form six valence pairs:
    • Intelligence/Stupidity
    • Calmness/Aggression
    • Sophistication/Unsophistication
    • Politeness/Rudeness
    • Articulation/Incoherence
    • Determination/Laziness
  • The traits are grounded in the Princeton Trilogy and social-psychology work on racial stereotypes.
  • The method builds on matched-guise probing (Hofmann et al.). It replaces free-form trait generation or log-likelihoods with closed-set Likert ratings, and it asks about the text's content rather than the speaker.
  • Prior work suggested otherwise, but here explicit dialect labels made bias larger than in the covert setting. Overt bias persists even in safety-aligned models.
  • In the finetuning approach, the model's own absolute-mode SAE score is the target for both members of each pair. For LLaMA-3.1-8B this reduced anti-AAVE gaps on several traits, including Intelligence, Politeness, Sophistication and Articulation.
  • A prompting-based debiasing baseline often helped. It was unreliable, however, because results depended on prompt wording and varied across sampling runs.
  • To reduce prompt sensitivity, each query is run with small perturbations, and the answers are combined by majority vote.

Interesting ideas

  • I1. The same item can be scored alone and then next to its counterfactual twin, which exposes biases that only appear when a system compares; this could serve as a cheap stress test for any judging component.
  • I2. Intent-equivalent pairs that differ only in dialect turn fairness into a measurable consistency check: the outputs for the two versions should match.
  • I3. Adding an explicit label (here, the dialect name) to an otherwise identical prompt works as a control condition, showing how much a surface cue alone shifts the output.
  • I4. Counterfactual finetuning uses the model's own output on the reference variant as the target for both variants, so no new human annotation is needed.
  • I5. Running small prompt perturbations and taking a majority vote makes LLM-based judgments more stable across runs.

For Mova

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

  • I1 (answer): Readers that choose between candidate answers could be checked with this test. Score each option both in isolation and in a paired line-up, and flag cases where the paired ranking flips or the gap widens.
  • I2 (tree): SAE/AAVE paired sentences could probe the perceptron parser and gparse. The check is whether the core predicate-argument edges of the UD tree stay the same, with dialect conversion judged by that agreement.
  • I3 (learn): LLM teacher tiers could be run with and without metadata cues as a negative control. Silver labels from teachers whose output moves with irrelevant labels would be rejected.
  • I4 (learn): Frames or tree outputs from a paraphrase's canonical form could be reused as silver targets for its variant. This trains invariance cheaply, with the authors' caveat that it privileges one reference form.
  • I5 (learn): The same perturb-and-vote scheme could combine LLM teacher answers before they enter silver data. This would cut noise caused by wording sensitivity.

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

🎓University of Southern CaliforniaLos Angeles · United States162 works · 35,836 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel, Ogheneyoma Akoni, Jevon Torres, Jaspreet Ranjit, Matthew Finlayson and Swabha Swayamdipta. 2026. Side-by-side Comparison Amplifies Dialect Bias in Language Models. arXiv:2605.24384 · In The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1145/3805689.3812217 · arXiv 2605.24384

@misc{kondapally2026side,
  title = {{Side-by-side Comparison Amplifies Dialect Bias in Language Models}},
  author = {Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel, Ogheneyoma Akoni, Jevon Torres, Jaspreet Ranjit, Matthew Finlayson and Swabha Swayamdipta},
  year = {2026},
  doi = {10.1145/3805689.3812217},
  eprint = {2605.24384},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1145/3805689.3812217}
}

References

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

  1. 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). 386–397. find
  2. Jiafu An, Difang Huang, Chen Lin, and Mingzhu Tai. 2025. Measuring gender and racial biases in large language models: Intersectional evidence from automated resume evaluation. PNAS nexus 4, 3 (2025), pgaf089. ↗
  3. Lena Armstrong, Abbey Liu, Stephen MacNeil, and Danaë Metaxa. 2024. The silicon ceiling: Auditing gpt’s race and gender biases in hiring. In Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. 1–18. find
  4. April Baker-Bell. 2019. Dismantling anti-black linguistic racism in English language arts classrooms: Toward an anti-racist black language pedagogy. Theory Into Practice 59 (10 2019). doi:10.1080/00405841.2019.1665415 ↗
  5. Peter Ball. 1983. Stereotypes of Anglo-Saxon and non-Anglo-Saxon accents: Some exploratory Australian studies with the matched guise technique. Language sciences 5, 2 (1983), 163–183. Side-by-side Comparison Amplifies Dialect Bias in Language Models FAccT ’26, June 25–28, 2026, Montreal, QC, Canada ↗
  6. 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 ↗
  7. J Stewart Black and Patrick van Esch. 2020. AI-enabled recruiting: What is it and how should a manager use it? Business horizons 63, 2 (2020), 215–226. ↗
  8. Su Lin Blodgett, Lisa Green, and Brendan O’Connor. 2016. Demographic dialectal variation in social media: A case study of African-American English. In Proceedings of the 2016 conference on empirical methods in natural language processing. 1119–1130.
  9. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems 29 (2016). ↗
  10. Minh Duc Bui, Carolin Holtermann, Valentin Hofmann, Anne Lauscher, and Katharina von der Wense. 2025. Large Language Models Discriminate Against Speakers of German Dialects. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (Eds.). Association for Computational Linguistics, Suzhou, China, 8212–8240. doi:10.18653/v1/2025.emnlp-main.415 ↗
  11. Jiali Cheng and Hadi Amiri. 2025. Linguistic Blind Spots of Large Language Models. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics. 1–17. ↗
  12. Anna Chung. 2019. How Automated Tools Discriminate Against Black Language. https://civic.mit.edu/index.html%3Fp=2402.html Civic Media. ↗
  13. Jacob Cohen. 2013. Statistical power analysis for the behavioral sciences. routledge. ↗
  14. DeepSeek. 2025. DeepSeek-V3. https://api-docs.deepseek.com/ Accessed: 2025-05-05. ↗
  15. Eve Fleisig, Genevieve Smith, Madeline Bossi, Ishita Rustagi, Xavier Yin, and Dan Klein. 2024. Linguistic bias in ChatGPT: Language models reinforce dialect discrimination. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 13541–13564.
  16. Arturo Fredes and Jordi Vitria. 2024. Using llms for explaining sets of counterfactual examples to final users. arXiv preprint arXiv:2408.15133 (2024). ↗
  17. Sahaj Garg, Vincent Perot, Nicole Limtiaco, Ankur Taly, Ed H. Chi, and Alex Beutel. 2019. Counterfactual Fairness in Text Classification through Robustness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (Honolulu, HI, USA) (AIES ’19). Association for Computing Machinery, New York, NY, USA, 219–226. doi:10.1145/3306618.3317950 ↗
  18. Gustave M Gilbert. 1951. Stereotype persistence and change among college students. The Journal of Abnormal and Social Psychology 46, 2 (1951), 245. ↗
  19. 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. 2796–2814. ↗
  20. Sophie Groenwold, Lily Ou, Aesha Parekh, Samhita Honnavalli, Sharon Levy, Diba Mirza, and William Yang Wang. 2020. Investigating African-American Vernacular English in transformer-based text generation. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP). 5877–5883.
  21. Yufei Guo, Muzhe Guo, Juntao Su, Zhou Yang, Mengqiu Zhu, Hongfei Li, Mengyang Qiu, and Shuo Shuo Liu. 2024. Bias in large language models: Origin, evaluation, and mitigation. arXiv preprint arXiv:2411.10915 (2024). ↗
  22. Abhay Gupta, Ece Yurtseven, Philip Meng, and Kevin Zhu. 2024. Aavenue: Detecting llm biases on nlu tasks in aave via a novel benchmark. In Proceedings of the Third Workshop on NLP for Positive Impact. 327–333.
  23. 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 (2024), 147–154. ↗
  24. Hawon Jeong, ChaeHun Park, Jimin Hong, Hojoon Lee, and Jaegul Choo. 2025. The Comparative Trap: Pairwise Comparisons Amplifies Biased Preferences of LLM Evaluators. In Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 79–108. find
  25. Marvin Karlins, Thomas L Coffman, and Gary Walters. 1969. On the fading of social stereotypes: Studies in three generations of college students. Journal of personality and social psychology 13, 1 (1969), 1. ↗
  26. Daniel Katz and Kenneth Braly. 1933. Racial stereotypes of one hundred college students. The Journal of Abnormal and Social Psychology 28, 3 (1933), 280. ↗
  27. Woojin Kim and Hyeoncheol Kim. 2025. Counterfactual fairness evaluation of machine learning models on educational datasets. In International Conference on Intelligent Tutoring Systems. Springer, 88–103. find
  28. Courtney A Kurinec and Charles A Weaver III. 2021. “Sounding Black”: Speech stereotypicality activates racial stereotypes and expectations about appearance. Frontiers in psychology 12 (2021), 785283. find
  29. Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. 2017. Counterfactual fairness. Advances in neural information processing systems 30 (2017). FAccT ’26, June 25–28, 2026, Montreal, QC, Canada Kondapally et al. find
  30. Wallace E Lambert, Richard C Hodgson, Robert C Gardner, and Samuel Fillenbaum. 1960. Evaluational reactions to spoken languages. The journal of abnormal and social psychology 60, 1 (1960), 44. ↗
  31. Jonathan Lazar, Jinjuan Heidi Feng, and Harry Hochheiser. 2017. Research methods in human-computer interaction. Morgan Kaufmann. find
  32. Sharon Gabriel Levy. 2023. Responsible AI via responsible large language models. University of California, Santa Barbara. find
  33. Masha Medvedeva, Michel Vols, and Martijn Wieling. 2020. Using machine learning to predict decisions of the European Court of Human Rights. Artif. Intell. Law 28, 2 (June 2020), 237–266. doi:10.1007/s10506-019-09255-y ↗
  34. Meta. 2024. Llama-3.1-8B. https://huggingface.co/meta-llama/Llama-3.1-8B Accessed: 2025-05-05. ↗
  35. OpenAI. 2023. GPT 3.5-Turbo. https://developers.openai.com/api/docs/models/gpt-3.5-turbo Accessed: 2025-05-05. ↗
  36. Kay Payne, Joe Downing, and John Christopher Fleming. 2000. Speaking Ebonics in a professional context: The role of ethos/source credibility and perceived sociability of the speaker. Journal of technical writing and communication 30, 4 (2000), 367–383. ↗
  37. Manish Raghavan, Solon Barocas, Jon Kleinberg, and Karen Levy. 2020. Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 conference on fairness, accountability, and transparency. 469–481. ↗
  38. Alvin Rajkomar, Michaela Hardt, Michael D Howell, Greg Corrado, and Marshall H Chin. 2018. Ensuring fairness in machine learning to advance health equity. Annals of internal medicine 169, 12 (2018), 866–872. ↗
  39. John R Rickford. 2016. Raciolinguistics: How language shapes our ideas about race. Oxford University Press. find
  40. Mihaela Rotar, Theresia Veronika Rampisela, and Maria Maistro. 2026. Can Fairness Be Prompted? Prompt-Based Debiasing Strategies in High-Stakes Recommendations. arXiv preprint arXiv:2603.12935 (2026). ↗
  41. Eleanor Shearer, S. Martin, A. Petheram, and R. Stirling. 2019. Racial Bias in Natural Language Processing. Oxford Insights. https://oxfordinsights.com/wp-content/uploads/2024/07/SHARED_-Racial-Bias-in-Natural-Language-Processing.pdf ↗
  42. Siqi Shen, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, Soujanya Poria, and Rada Mihalcea. 2024. Understanding the capabilities and limitations of large language models for cultural commonsense. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 5668–5680. ↗
  43. Amir Taubenfeld, Tom Sheffer, Eran Ofek, Amir Feder, Ariel Goldstein, Zorik Gekhman, and Gal Yona. 2025. Confidence improves self-consistency in llms. In Findings of the Association for Computational Linguistics: ACL 2025. 20090–20111. find
  44. Guangya Wan, Yuqi Wu, Jie Chen, and Sheng Li. 2025. Reasoning aware self-consistency: Leveraging reasoning paths for efficient llm sampling. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 3613–3635. ↗
  45. Shen Wang, Tianlong Xu, Hang Li, Chaoli Zhang, Joleen Liang, Jiliang Tang, Philip S. Yu, and Qingsong Wen. 2024. Large Language Models for Education: A survey and outlook. IEEE Signal Processing Magazine 42 (2024), 51–63. https://api.semanticscholar.org/CorpusID: 268723753 ↗
  46. Sade Wilson. 2012. African American English: Dialect mistaken as an articulation disorder. McNair Scholars Research Journal 4, 1 (2012), 11. find
  47. Tian Xie, Tongxin Yin, Vaishakh Keshava, Xueru Zhang, and Siddhartha Reddy Jonnalagadda. 2025. Biascause: Evaluate socially biased causal reasoning of large language models. arXiv preprint arXiv:2504.07997 (2025). ↗
  48. Runtao Zhou, Guangya Wan, Saadia Gabriel, Sheng Li, Alexander Gates, Maarten Sap, and Thomas Hartvigsen. 2025. Disparities in LLM Reasoning Accuracy and Explanations: A Case Study on African American English. (03 2025). doi:10.48550/arXiv.2503.04099 Side-by-side Comparison Amplifies Dialect Bias in Language Models A FAccT ’26, June 25–28, 2026, Montreal, QC, Canada Score Frequency Dominance Patterns ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kondapally-2026-side-side-comparison-amplifies-dialect/