Speakers of non-“standard” varieties of English face discrimination on the basis of how they speak, an issue that language models can exacerbate. We present a large-scale study of linguistic bias exhibited by GPT-3.5 Turbo and GPT-4 covering ten different dialects of English, including eight widely spoken non-“standard” varieties from around the world. We prompted the models with text by native speakers of each variety and analyzed the responses via detailed linguistic feature annotation and native speaker evaluation. We find that the models default to and perform best for standard varieties of English. Across the other varieties, we find that the models consistently exhibit issues including lack of comprehension, stereotyping and demeaning content, and condescending responses. We also find that if these models are asked to imitate the writing style of prompts in non-standard varieties, they produce text that is more unnatural, condescending, demeaning, and especially prone to stereotyping. These results suggest that language models exhibit lower quality of service for speakers of these varieties and exacerbate harms faced due to linguistic discrimination.
Eve Fleisig is a third-year PhD student at UC Berkeley advised by Dan Klein. Her research lies at the intersection of natural language processing and AI ethics, with a focus on preventing societal harms of language models and ensuring that AI systems account for the perspectives of diverse populations. Previously, she received a B.S. in computer science from Princeton University. Her research has been awarded an NSF Graduate Research Fellowship, Berkeley Chancellor’s Fellowship, and EMNLP Outstanding Paper Award.