AI models rank English accents differently, raising fairness concerns
Leading AI models tend to give higher scores to speakers with American and British accents than to those speaking English with Chinese, Nigerian or Indian accents, even when they deliver identical content, a new study by the University of Washington has found, Qazinform News Agency correspondent reports.
The researchers tested four AI models across three settings: job interviews, conference presentations and spoken English proficiency tests. The models included Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio 1.5 and the open-source Qwen 3 Omni.
The systems evaluated speech from 5 accent groups: American, British, Chinese, Indian and Nigerian. They rated both delivery and content on a 7-point scale.
American and British speakers received the highest scores, Indian speakers ranked in the middle, while Chinese and Nigerian speakers generally received the lowest ratings.
Importantly, the models could accurately transcribe the accented speech. However, they still gave lower scores to some speakers, suggesting that the differences were not simply due to difficulties understanding their words.
The researchers also tested the models with recordings from 52 human participants representing the five countries. The differences in scores between human speakers were even greater than those observed with AI-generated voices.
The study found that accents acoustically further from American English tended to receive lower ratings. The models' explanations also described some Chinese and Nigerian speakers as difficult to understand, despite successfully transcribing their speech.
Researchers said such bias could originate in training data that underrepresents certain accents or portrays accented speakers as difficult to understand. Training methods that reinforce conventional expectations of professional speech could also contribute.
The findings raise concerns as AI tools increasingly assess people in situations where fair evaluations matter, including recruitment, education and English proficiency tests used in immigration procedures.
Senior author Yulia Tsvetkov, an associate professor at the university, said developers should build more representative training and evaluation datasets. She also called for testing model performance across different populations, dialects and accents rather than relying only on average accuracy.
Earlier, Qazinform News Agency reported on how language models, such as GPT-5 reflect political bias and emotional cues.
