Indian accent faces hidden bias in AI-powered job interviews and English tests, University of Washington study reveals. (AI Image) Artificial intelligence models may favour American and British English accents over Indian English when evaluating job interviews, presentations and spoken language tests, according to a study by researchers at the University of Washington.
The research found that leading AI models generally gave higher ratings to speakers with American and British accents, while Indian speakers received scores in the middle, ahead of Chinese and Nigerian speakers.The study, titled The Cost of Sounding Different: Accent Bias in Audio Language Models, examined four AI models and found differences in how they assessed speech delivery and content, even when speakers communicated identical information.
The findings raise questions about the use of AI-based assessment tools in recruitment, education and English proficiency testing, particularly when candidates come from different linguistic backgrounds.University of Washington study examines five English accents The researchers assessed how AI models evaluated English speech across five accents: American, British, Indian, Chinese and Nigerian.
The assessment covered three areas: job interviews, presentations and spoken English proficiency tests.
The team tested four AI models: Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio 1.5 and the open-source Qwen 3 Omni.To conduct the research at scale, the team used AI-generated voices from ElevenLabs to read identical scripts in different accents.
These included interview responses, classroom presentations, conference presentations and English test responses.
Both male and female voices were included.The models rated speakers on a scale of one to seven for delivery and content, while also providing explanations for their scores.
Researchers subsequently validated the findings using 52 human speakers from the five countries.Indian English ranked in the middle, American and British accents scored higher The findings showed a consistent preference for American and British accents across the four models.
Chinese and Nigerian speakers generally received the lowest ratings, while Indian speakers were placed in the middle.The researchers also found that the difference was not limited to speech delivery.
Even when the spoken content was identical and the models could accurately transcribe the speech, they rated both the delivery and content of American and British speakers more favourably.According to the University of Washington, the research team also examined the acoustic characteristics of different accents and found that accents with a greater acoustic distance from American English were often rated lower.AI models understood accented speech but rated speakers differently An important finding was that the models could accurately transcribe accented English, despite giving some speakers lower evaluation scores."Even though they repeatedly claimed Chinese and Nigerian speakers 'mumbled' or were 'difficult to understand,' the models were actually very good at transcribing the accented speech," said Orevaoghene Ahia, co-lead author and a doctoral student at the University of Washington's Paul G.
Allen School of Computer Science & Engineering.Ahia added that the models understood the speakers but appeared to have an implicit hierarchy based on accents.Speaking about the implications of the findings, Ahia said people were increasingly relying on audio AI models as third-party evaluators in hiring, oral examinations and English tests used in immigration processes, where fairness was particularly important.Join conversation Share your thoughts in the comments Be respectful · TOI community guidelines Researchers flag implications for AI-based assessments Senior author Yulia Tsvetkov, an associate professor at the University of Washington's Allen School, said the team selected leading AI models because these were the systems companies were likely to deploy in real-world settings.Tsvetkov said one direction for future research was to develop more representative training and evaluation datasets to improve the accuracy and fairness of language models.The research team also plans to expand its work to include more human accents, including regional variations within countries, and examine how AI models evaluate audio in medical settings, such as automated emergency responses.The study was presented on October 25, 2026, at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary.