Study: Accent drives audio AI's assessments in job interviews and language tests

A new study from the University of Washington shows that four of the most advanced audio models rate American and British accents highest and Chinese and Nigerian accents lowest – even when all speakers read exactly the same script and the…

Illustration: A row of identical tuning forks, two bent and pushed into shadow — a metaphor for audio AI judging identical voices differently by accent.
Illustration

Study: Accent drives audio AI's assessments in job interviews and language tests

A new study from the University of Washington shows that four of the most advanced audio models rate American and British accents highest and Chinese and Nigerian accents lowest – even when all speakers read exactly the same script and the models transcribe them flawlessly. The bias, in other words, lies not in comprehension but in judgment.

What the study shows

Researchers at the University of Washington tested how some of the most advanced AI models evaluate English speakers with different accents – both what they said and how they said it. The result, presented in October at EMNLP 2026 (the Conference on Empirical Methods in Natural Language Processing), is unambiguous: American and British accents ranked highest, Indian accents in the middle, and Chinese and Nigerian accents lowest, according to findings reported by The Seattle Times (bf3dd1e9).

What makes the finding far-reaching is that the bias cannot be explained by the models struggling to understand the speakers. In the scripted tests, all speakers said exactly the same words, and the models transcribed the accented speech correctly. The models' comprehension was fine – yet their assessments still differed systematically by accent.

How the models were tested

The researchers used AI-generated "synthetic voices" that read identical scripts in five different accents. The scripts mimicked three realistic genres: answers in job interviews, conference presentations, and answers on English-language tests. The models were asked to rate the speeches on a scale of one to seven on two dimensions: delivery and content. The results were then confirmed with human speakers (bf3dd1e9).

The models tested were Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio-1.5, and Qwen3-Omni from China's Alibaba. When the team ran the tests in the spring, they were among the most advanced models available. Newer versions have since been released, so it remains unclear whether current versions show the same pattern (bf3dd1e9).

One detail illustrates how widespread the pattern is: even Alibaba's model – a Chinese company's – gave British speakers the best delivery scores and Chinese speakers the worst when listening to humans answering job interview questions (bf3dd1e9).

When the models' explanations can't be trusted

The models justified their ratings in writing, and the justifications followed the same skewed pattern. They described American and British speakers as "clear" and "excellent," while Chinese and Nigerian speakers were described as "mumbling" or "difficult to understand" (bf3dd1e9) – in contrast to the transcripts, which showed the models understood them just fine.

Even more problematic is the researchers' finding on self-reporting: the models admitted that accent had influenced their assessment in only roughly 10 to 30 percent of explanations. The research therefore concludes that AI cannot be trusted to self-report its own bias (bf3dd1e9). In other words: a model that uses accent as a hidden evaluation criterion will usually offer a justification that pretends the criterion was something else.

Why it matters

"People now rely on audio AI models as a kind of third-party evaluator – for hiring, for oral exams in education, for certain English tests in immigration contexts – settings where fairness really matters," said Orevaoghene Ahia, a doctoral student at the Paul G. Allen School of Computer Science and Engineering in Seattle and a co-author of the study (bf3dd1e9).

The South China Morning Post, which carried the study via Tribune News Service, points to the same implications: the models rated the content of identical scripts differently, with consequences for the use of AI in hiring, education, and immigration testing (ee4f931f).

It is the combination of the two findings that makes the verdict serious for these applications: the model judges unequally by accent, and it will usually not admit it if you ask it why.

Caveats and what remains

The findings are known through secondary coverage from The Seattle Times and SCMP, not through primary reading of the study itself. The sources do not state the number of speakers, statistical significance, or the full author list, and the exact date of the October EMNLP presentation is not specified. That is worth keeping in mind when weighing how robust the findings are.

The funding is also worth noting: gifts from Google and OpenAI partially financed the research – both companies had models in the study. Ahia stated, however, that the companies played no role in the work (bf3dd1e9).

The researchers also found no meaningful difference between male and female voices, according to Ahia. Going forward, the team plans to study more accents and examine audio AI in emergency dispatch centers such as 911 call centers – a context where a misunderstanding affects not just fair evaluation but potentially life and health (bf3dd1e9).

The sources say nothing about why the models carry this accent bias, and it remains unclear whether newer model versions show the same pattern. What the researchers can document, however, is clear enough: when audio AI is used as a judge, the accent determines how the same words are judged – and the model won't tell you itself.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.