A University of Washington-led study has found that leading AI models favour American and British accents when evaluating spoken English, potentially disadvantaging speakers from other linguistic backgrounds in hiring and academic assessments.
Researchers at the University of Washington (UW) have found that state-of-the-art artificial intelligence models display bias against certain English accents when assessing job interviews, presentations and spoken English proficiency tests.
The study examined four AI models, including Google’s Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI’s GPT-Audio 1.5 and the open-source Qwen 3 Omni. Researchers found that the models generally gave the lowest scores to Chinese and Nigerian accents, placed Indian accents in the middle, and ranked American and British accents highest.
This pattern persisted even when speakers delivered identical content and the AI systems could accurately transcribe their speech. The models consistently gave higher ratings for both delivery and content to speakers with American and British accents.
The research was presented on October 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary.
“People are now trusting audio AI models as sort of third-party evaluators, for hiring, for oral exams in education, for certain English tests in immigration — contexts where fairness really matters,” said co-lead author Orevaoghene Ahia, a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering. “So I was curious how models perceive speech from people who don’t speak with so-called ‘standard’ American accents.”
Researchers Tested Four AI Models Across Three Domains
The research team selected English accents from five regions with large English-speaking populations: American, British, Chinese, Indian and Nigerian. The researchers examined how AI models assessed speakers across three settings: job interviews, presentations and English proficiency tests.
To conduct the experiments at scale, the team used AI-generated voices from ElevenLabs to read identical scripts in different accents. The scripts included interview answers, classroom and conference presentations, and responses to English proficiency tests. Both male and female voices were included.
The four AI models evaluated each response on a scale of one to seven, assigning separate scores for delivery and content. They also provided brief explanations for their assessments.
Researchers subsequently tested the interview responses with 52 human participants from the five countries. Each participant read four scripted answers and also responded to questions without a prepared script.
The results again favoured American and British accents. Researchers also found that the difference in scores between human speakers with different accents was greater than the disparity between the AI-generated voices.
Senior author Yulia Tsvetkov, a UW associate professor in the Allen School, explained why the team selected these particular systems.
“We wanted to check these state-of-the-art models,” Tsvetkov said, “because these are what companies would deploy in a real-world setting.”
Acoustic Differences May Influence AI Assessments
To understand the scoring disparities, the researchers examined the acoustic characteristics of the different accents. They measured what they described as their “acoustic distance” and found that accents further from American English were frequently assigned lower scores.
The bias was also apparent in the models’ explanations. Although the systems sometimes described Chinese and Nigerian speakers as mumbling or difficult to understand, they were generally able to transcribe the accented speech accurately.
“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,” Ahia said. “They understood the speakers, but they just had this implicit notion of hierarchy among them based on accent.”
Researchers Call for More Representative AI Training Data
The study suggests that accent bias can enter AI models through several possible routes. These include training data that characterises accented speakers as difficult to understand, datasets that inadequately represent certain populations and accents, and training methods that reinforce conventional expectations of professional speech.
The researchers highlighted that English is an official language in Nigeria and is used as the language of instruction in Indian schools. Nevertheless, Nigerian and Indian accents received lower ratings than American and British accents.
“One clear direction for future work is to create more representative training and evaluation data sets that could improve the accuracy and fairness of language models,” Tsvetkov said.
The team intends to extend the research by examining a wider range of human accents, including regional variations within individual countries. It also plans to investigate how AI models evaluate audio in medical settings, including automated emergency response.
Tsvetkov emphasised that developers should assess model performance across different population groups and dialects rather than relying only on overall accuracy.
“For all of these, we hope that developers build and publish evaluations of their models that don’t focus only on average performance, but also break it down into subpopulations, dialects and so on,” Tsvetkov said. “We don’t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%.”
Read More :