Picture two job candidates, sitting in the same virtual interview room, giving the same answer to the same question, word for word. Their qualifications are identical. Their tone is steady. Their words are perfectly rehearsed. The only difference is the voice that delivers them: one has an American accent, the other a Nigerian accent. Ask a leading artificial intelligence model to score those performances, and the American voice will almost certainly come out ahead. That is the core finding of a sobering new study from the University of Washington, which set out to understand how today’s most powerful AI systems judge the way people speak. The researchers tested models from Google, OpenAI, and Alibaba, asking them to evaluate spoken English in five accents: American, British, Chinese, Indian, and Nigerian. Across simulated job interviews, conference presentations, and English proficiency tests, the pattern was remarkably consistent. American and British speakers were ranked highest, Indian speakers fell somewhere in the middle, and Chinese and Nigerian speakers were consistently ranked lowest. The algorithms did not hear gibberish. They heard perfectly intelligible English. But their scoring told a different story, one that has serious consequences for anyone whose accent does not fit a narrow, American-centric mold.
What makes this bias especially disturbing is how subtle and almost invisible it is. The researchers went out of their way to isolate the variable: every speaker read the exact same script, and the AI models were explicitly instructed to judge only the substance of the answer, not the style of delivery. Yet the penalty still fell mostly on delivery, with Chinese and Nigerian speakers marked down for their pronunciation, pacing, and perceived clarity. Even worse, the content scores for these same speakers sometimes slipped as well, even though every word they said was identical to what the American and British speakers said. The models seemed unable to separate the message from the messenger. Orevaoghene Ahia, a UW doctoral student and co-lead author of the study, noted a strange contradiction: “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.” In other words, the AI understood every word perfectly, but still decided that the way it sounded was worse. It is as if a human listener heard you clearly, nodded along, and then gave you a lower grade because of the way you talked, while insisting the grade had nothing to do with your accent. That may sound like a failure of common sense, but for an algorithm, it is simply a reflection of what it has been taught.
The real-world stakes of this go far beyond a low score on some research exercise. Companies are increasingly using AI to screen recorded job interviews, sorting through thousands of candidates before a human ever looks at a résumé. Instructors are turning to oral exams as written work becomes easier to fake with AI, hoping that spoken assessments can reveal what a student actually knows. Perhaps most consequential of all, at least one major English test used for visas and immigration, Pearson’s PTE Academic, relies on AI to score spoken answers. That test was taken about 1.2 million times in 2023, according to figures cited in the study. For millions of people around the world, a visa, a job offer, or a place at a university can hinge on an algorithm’s judgment of their accent. The researchers were careful to note that they did not test the actual products used by hiring platforms or test makers, because those systems are usually closed and proprietary, and there are no public audits of AI-scored English tests that the authors could find. Instead, they tested general-purpose models, the kind of technology that companies would likely build on to create their own tools. “We wanted to check these state-of-the-art models because these are what companies would deploy in a real-world setting,” said Yulia Tsvetkov, a UW associate professor and the study’s senior author. That is a crucial point: the bias is not hidden in some obscure algorithm, but baked into the very models that shape the future of automated decision-making.
So why does this happen? The researchers looked at the acoustics, measuring how different each accent was from American English. The results were strikingly linear: Chinese and Nigerian accents were the farthest away from the American baseline, and the models’ scores dropped as that distance grew. Indian and British accents sat much closer to the American norm, and their scores reflected that proximity. This suggests that the problem is not about clarity or correctness, but about familiarity. The models were trained on enormous amounts of data, and that data appears to be saturated with certain assumptions about how English should sound. The researchers suggest that bias may come from training data that explicitly or implicitly describes accented speakers as unclear, from datasets that underrepresent certain accents, or from training processes that reinforce narrow ideas about what a professional voice should be. What makes this even more troubling is that English is an official language in Nigeria and the language of instruction in Indian schools, yet both accents were consistently rated below American and British ones. This is not about actual proficiency. It is about a learned hierarchy of voices, one that places a particular kind of speaker at the top and everyone else somewhere below. The algorithm is not evaluating English; it is evaluating identity, and doing so with the cold confidence of a system that believes it is being objective.
This study is not an isolated warning. It is part of a broader pattern of AI bias that researchers at the University of Washington have been documenting for years. Earlier projects found that AI models favored résumés with white- and male-associated names, meaning a qualified woman or person of color could be passed over before anyone even looked at their experience. Other research showed that politically slanted chatbots could quietly shift people’s views, raising concerns about how these tools are used in public discourse. And in a particularly telling example, AI-generated children’s stories systematically erased female animal characters, replacing them with males in ways that reinforced stereotypes. Each of these findings points to the same unsettling truth: AI does not need to be deliberately taught to discriminate. It absorbs discrimination from the data it is trained on, then reproduces it at scale. The new accent study, presented on Oct. 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest, adds a voice-based dimension to this growing body of evidence. The researchers plan to continue this line of work, expanding to more accents, including regional variations within countries, and exploring how AI handles audio in high-stakes settings like automated emergency response. Imagine calling for help in a moment of crisis and being understood perfectly, yet somehow being taken less seriously because of the way you pronounce a word. That is not a distant dystopia; it is a plausible future if these biases are left unchecked.
The question is what we do about it. Tsvetkov and her colleagues are not calling for AI to be abandoned, but for it to be held to a higher standard. “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,” she said. “We don’t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%.” That is a powerful way to frame the issue. A model can look great on paper while quietly failing the very people who need it to work. If AI is going to sit in on job interviews, grade English exams, or answer emergency calls, we have a right to know who it is punishing and why. We have a right to demand that the data it learns from includes the full richness of human voices, not just the ones that fit a comfortable norm. And we have a right to insist that the people who build these systems are accountable for the biases they encode, whether intentionally or not. The promise of AI has always been that it could be more objective than humans, that it could rise above our prejudices and make decisions based purely on merit. But this study is a reminder that AI is a mirror, not a savior. It reflects the biases of the world it is trained on, and unless we are careful, it will amplify them. The first step toward fixing that is admitting the problem exists. Thanks to this research, we can no longer pretend that an accent is just a sound. For millions of people, it is the difference between being heard and being dismissed.
