Study Finds AI Models Penalize Chinese and Nigerian Accents

University of Washington researchers found that AI models often gave lower scores to English speakers with Chinese and Nigerian accents, according to GeekWire. The study tested models from Google, OpenAI and Alibaba across job interviews, conference presentations and English proficiency tests.

American and British speakers generally received the highest scores, followed by Indian speakers. Chinese and Nigerian speakers received the lowest scores, although every participant read the same scripts. Scores mainly differed on delivery, but the models sometimes lowered content scores despite instructions to assess only substance.

The models accurately transcribed the accented speech even when their evaluations described some speakers as difficult to understand. The researchers tested general purpose models rather than systems used by hiring platforms or test providers. They presented the study on 25 October at the Conference on Empirical Methods in Natural Language Processing in Budapest. The research was funded in part by gifts from Google and OpenAI, whose models were among those tested.