AI models favor some accents over others, researchers find
Published in Science & Technology News
Ask an artificial intelligence model to grade a job interview or test answers, and a person’s accent may change the score.
That’s according to new University of Washington research that looked at how some of the most advanced AI models rated English speakers with different accents, both on what they said and how they said it.
In the scripted tests, every speaker said the same words, and AI models transcribed the accented speech accurately, meaning the models understood the accents just fine. But the models still scored speakers differently, revealing a bias in how they rated accents.
American- and British-sounding speakers ranked highest, speakers with Indian accents were in the middle and those with Chinese and Nigerian accents scored lowest, according to the findings.
The AI models even rated the content of identical scripts differently, despite being told to judge only what was said.
“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 Orevaoghene Ahia, a UW doctoral student in the Paul G. Allen School of Computer Science and Engineering and a co-author of the study, which was presented in October at the 2026 Conference on Empirical Methods in Natural Language Processing.
“Caution is definitely necessary,” Ahia added.
The researchers tested Google’s Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI’s GPT-Audio-1.5 and Qwen3-Omni, made by the Chinese tech giant Alibaba. The models were among the most advanced available when the team ran its tests in the spring. Newer versions have since been released.
For part of the test, AI-generated “synthetic voices” read identical scripts with the five different accents, mimicking job-interview answers, conference presentations and English test responses.
The models listened and rated each recording from one to seven on delivery and content.
To make sure the results held up with natural speech, the researchers also had people with each accent deliver the same scripted interview answers. For both synthetic and human voices, the AI models ranked American and British accents highest.
Even Alibaba’s model, listening to people’s interview answers, gave British speakers the best delivery scores and Chinese speakers the worst.
Results didn’t meaningfully differ between men’s and women’s voices, Ahia said.
The models’ written explanations of the rankings said American and British speakers were “clear” and “excellent,” whereas Chinese and Nigerian speakers were “mumbling” or “difficult to understand.”
“They understood the speakers, but they just had this implicit notion of hierarchy among them based on accent,” Ahia said of the AI models.
The research also found AI cannot be trusted to self-report its bias. Models admitted that accents shaped ratings in only about 10% to 30% of explanations.
Gifts from Google and OpenAI partly funded the research, but Ahia said the companies had no role in the work.
While developers of AI models measure their accuracy, Ahia said this isn’t enough and that testing for fairness across accents should be a standard step.
The research team next wants to study more accents and how AI handles audio in emergency response scenarios, like 911 dispatch centers. Ahia said that work is still in its early stages because the data is sensitive. Local governments are already using audio AI tools on nonemergency and 911 dispatch lines.
©2026 The Seattle Times. Visit seattletimes.com. Distributed by Tribune Content Agency, LLC.







Comments