AI Models Penalize Non-Native English Speakers, UW Study Finds

Leading AI models from tech giants like Google and OpenAI exhibit significant accent bias, consistently penalizing non-native English speakers in simulated interviews and proficiency tests, according to groundbreaking research from the University of Washington. The study, which also included models from Alibaba, found that these systems generally scored American and British English speakers highest, while Chinese and Nigerian English speakers received the lowest evaluations, even when delivering identical content.

Published on November 2, 2026, the UW study revealed that this bias primarily impacted "delivery" scores, with AI models frequently claiming Chinese and Nigerian speakers "mumbled" or were "difficult to understand." Surprisingly, the researchers noted that despite these subjective criticisms, the models demonstrated high accuracy in transcribing the accented speech, indicating the issue wasn't a failure to comprehend the words themselves, but rather a negative judgment based on accent.

"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," stated Orevaoghene Ahia, a UW doctoral student and co-lead author. This disparity suggests a deeper issue where accent itself triggers a negative evaluation, rather than a genuine comprehension problem.

This accent bias carries substantial real-world implications. Companies increasingly rely on AI for initial job interview screenings, and educational institutions are turning to AI-scored oral exams as a countermeasure against AI-generated written work. Furthermore, major English proficiency tests, such as Pearson’s PTE Academic, which was administered approximately 1.2 million times in 2023, utilize AI for scoring spoken responses.

Although the UW team tested general-purpose AI models rather than proprietary systems used by specific hiring platforms or test providers, their findings are critical. These state-of-the-art models form the technological bedrock upon which companies build and deploy their specialized applications, meaning the biases identified could permeate various high-stakes scenarios.

Researchers observed a correlation between the acoustical distance of an accent from American English and the scores received. Chinese and Nigerian accents were found to be acoustically farthest from the American baseline, aligning with their lower scores, while Indian and British accents, acoustically closer, received higher rankings. This pattern suggests that the training data for these AI models might be implicitly or explicitly biased against certain vocal characteristics.

The underlying causes of this bias could stem from several factors. It might originate from training data that inaccurately describes accented speakers as unclear, from datasets that inadequately represent diverse accents, or from training processes that inadvertently reinforce preconceived notions about what a "professional" or "understandable" voice should sound like. This is particularly concerning given that English serves as an official language in Nigeria and the primary language of instruction in Indian schools, yet speakers of these accents consistently underperformed.

This research extends a series of prior UW studies uncovering various forms of AI bias. Earlier work demonstrated AI models favoring resumes with white- and male-associated names, chatbots swaying political opinions, and AI-generated children’s stories nearly eliminating female animal characters. These cumulative findings underscore a pervasive challenge in developing equitable AI systems.

Looking ahead, the UW researchers plan to broaden their investigation to include a wider array of accents, including regional variations within countries. They also intend to examine how AI models perform in critical audio contexts, such as automated emergency response systems in medical settings. The team emphasizes the need for developers to move beyond average performance metrics and publish detailed evaluations of their models across diverse subpopulations and dialects.

As Yulia Tsvetkov, a UW associate professor and senior author of the study, highlighted, "We don’t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%." This call to action underlines the urgency for more inclusive AI development practices to prevent technological advancements from exacerbating existing societal inequalities.

FAQ

Q: What exactly did the University of Washington study reveal about AI accent bias?

A: The study found that leading AI models from Google, OpenAI, and Alibaba consistently scored American and British English speakers higher than Chinese and Nigerian English speakers in simulated job interviews, presentations, and proficiency tests, even when all speakers recited identical content. This bias mainly affected perceived "delivery" quality, despite the models accurately transcribing all accents.

Q: What are the real-world consequences of this AI accent bias?

A: The bias could unfairly disadvantage individuals in critical applications such as AI-driven job screenings, oral academic exams, and standardized English proficiency tests required for visas or immigration. Speakers with certain accents might receive lower scores not due to their content or actual English proficiency, but solely because of their accent.

Q: How do researchers propose addressing this accent bias in AI models?

A: The UW team advocates for AI developers to conduct and publicly share comprehensive evaluations that assess model performance across diverse subpopulations, dialects, and accents, rather than focusing solely on overall average performance. Their ongoing research also includes studying more accents and exploring AI performance in critical fields like medical audio analysis.