A new study shows that large language models (LLMs) may exhibit stronger biases based on the dialect or linguistic style of a user than on explicit demographic information. The research, titled "Dialect vs Demographics," was covered in a recent episode of the Daily Papers AI podcast.
Researchers Irti Haq and Belén Saldías found that when LLMs process text written in non-standard dialects or with implicit linguistic signals, they can produce outputs that reflect harmful stereotypes or discriminatory assumptions. In contrast, when users explicitly state their demographic background, models often adjust to be more neutral.
"The model is picking up on subtle cues in language—like word choice or syntax—that are strongly correlated with certain groups, and then applying biased reasoning," the podcast hosts explained.
The findings highlight a challenge for AI fairness: even if developers remove explicit demographic markers from training data, biases can persist through language patterns. The paper suggests that current evaluation benchmarks may underestimate real-world bias because they typically use formal, standardized language.
The podcast episode discusses implications for chatbot design, content moderation, and inclusive AI. The full paper is available on arXiv.