AI Models Show Bias Against Certain Accents in Scoring Interviews and Tests, UW Researchers Find

In an increasingly AI-driven world, the fairness and impartiality of artificial intelligence models are under constant scrutiny. A recent study by a University of Washington (UW)-led research team has brought to light a significant concern: leading AI models demonstrate bias against certain accents when scoring spoken English in high-stakes situations. This reported development, presented at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary, on October 25, 2026, has important implications for students, job applicants, and anyone assessed by AI speech recognition tools.

According to UW News, the researchers investigated how state-of-the-art multimodal language models, including two from Google and one from OpenAI, evaluated human speech. Their findings indicate a consistent pattern where models generally ranked Chinese and Nigerian accents the lowest, Indian accents in the middle, and American and British accents the highest. This hierarchy persisted even when the spoken content was identical and accurately transcribed by the AI, suggesting a bias rooted in accent rather than comprehension.

The Scope of the Bias: From Hiring to Education

The UW team, led by doctoral student Orevaoghene Ahia and associate professor Yulia Tsvetkov from the Paul G. Allen School of Computer Science & Engineering, focused on three critical domains where AI-powered speech evaluation is becoming prevalent: job interviews, conference presentations, and spoken English proficiency tests. These are contexts where the stakes are high, directly impacting an individual's career prospects, academic opportunities, or immigration status.

According to Ahia, the motivation behind the study was to understand how these AI models, which are increasingly trusted as “third-party evaluators,” perceive speech from individuals who do not speak with what might be considered “standard” American accents. The researchers tested responses from voices speaking English with five different accents: American, British, Chinese, Indian, and Nigerian. To ensure consistency, they initially used AI voices from ElevenLabs to read identical texts, then validated these results with 52 human speakers from the five countries.

The study specifically examined four AI models: Google’s Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI’s GPT-Audio 1.5, and an open-source model, Qwen 3 Omni. Senior author Yulia Tsvetkov noted that these models were chosen because they represent the kind of advanced AI that companies are likely to deploy in real-world settings. The models were tasked with rating responses on a scale of one to seven for both delivery and content, providing explanations for their scores.

Understanding the Nature of the Bias

What makes these findings particularly concerning is that the bias appeared to be independent of the AI's ability to understand the spoken words. Ahia explained that despite the models frequently claiming Chinese and Nigerian speakers “mumbled” or were “difficult to understand,” the AI was often very good at transcribing the accented speech. “They understood the speakers, but they just had this implicit notion of hierarchy among them based on accent,” Ahia stated, according to UW News.

This suggests that the bias isn't merely a transcription error but a deeper issue within the model's evaluative framework. The research team analyzed features of the accents and measured their “acoustic distance,” finding that accents acoustically further from American English were generally ranked lower. This hierarchy was also evident in the AI models' explanations for their scores.

The UW researchers suggest that this bias can be introduced during the AI models' training phase. Potential sources include training data that describes accented speakers as difficult to understand, data sets that underrepresent certain populations and accents, or training methods that reinforce conventional expectations about what a “professional” speaker should sound like. It's important to remember that English is an official language in many countries, such as Nigeria, and a primary language of instruction in others, like India, yet these accents were rated lower than British and American ones in the study.

Implications for Students and Educators

For students, particularly those in K-12 and higher education, these findings underscore the importance of diverse and equitable learning environments. As AI tools become more integrated into educational assessments, such as evaluating oral presentations, language proficiency, or even participation, it's crucial to ensure these tools do not inadvertently penalize students based on their accent. Educators might consider these findings when selecting AI-powered assessment tools or when interpreting results from such systems.

Platforms like COSMIQ, a free voice-driven AI tutor, are designed to support learning without judgment, offering free voice AI tutoring to every K-12 student, free forever. Such tools can help learners practice and build confidence in their communication skills in a neutral environment, offering support for a wide range of subjects and even exam preparation, without the potential for accent-based bias in their core tutoring functions.

Moving Forward: Towards More Equitable AI

The UW team's research highlights a critical area for improvement in AI development. Tsvetkov emphasized that a clear direction for future work is to create more representative training and evaluation data sets to enhance the accuracy and fairness of language models. The team plans to expand their research to include more human accents, including regional variations, and to examine AI audio analysis in other sensitive contexts, such as medical applications.

Ultimately, the researchers advocate for developers to build and publish evaluations of their models that go beyond average performance, breaking down results by subpopulations, dialects, and other demographic factors. This approach would help prevent situations where a model might show high average accuracy but perform poorly for specific groups, perpetuating inequity. The work of the UW team, supported in part by gifts from Google and OpenAI, serves as a vital reminder that as AI becomes more pervasive, vigilance against bias and a commitment to fairness must remain at the forefront of its development and deployment.