{"id":957564,"date":"2026-11-04T04:10:14","date_gmt":"2026-11-04T04:10:14","guid":{"rendered":"https:\/\/www.newsbeep.com\/ca\/957564\/"},"modified":"2026-11-04T04:10:14","modified_gmt":"2026-11-04T04:10:14","slug":"leading-ai-models-are-biased-against-certain-accents-when-scoring-interviews-presentations-and-tests","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/ca\/957564\/","title":{"rendered":"Leading AI Models Are Biased Against Certain Accents When Scoring Interviews, Presentations and Tests"},"content":{"rendered":"

Newswise \u2014 Artificial intelligence models replicate, and sometimes amplify, all sorts of human biases. In the last few years, University of Washington researchers have found AI has erased female characters from kids\u2019 stories<\/a>, favored job applicants of certain races and genders<\/a>, and swayed human decisions<\/a>.\u00a0<\/p>\n

Much bias research has focused on text and visuals. Yet people speak to AI models with increasing regularity, and AI tools are now used to screen job applicants and score oral presentations. Now, a UW-led team of researchers found that state-of-the-art AI models were biased against speakers with certain accents.\u00a0<\/p>\n

The team investigated how state-of-the-art multimodal language models, including two from Google and one from OpenAI, evaluated human speech in job interviews, conference presentations and spoken English proficiency tests. The models scored responses from voices speaking English with five different accents: American, British, Chinese, Indian and Nigerian. The AI models generally ranked Chinese and Nigerian accents the lowest, Indian in the middle, and American and British the highest. Even when the content was exactly the same and the models could accurately transcribe the speech, the systems ranked both the delivery and content higher for the American and British speakers.\u00a0<\/p>\n

The team presented its research<\/a> Oct. 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary.\u00a0<\/p>\n

\u201cPeople 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 \u2014 contexts where fairness really matters,\u201d said co-lead author Orevaoghene Ahia<\/a>, a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering. \u201cSo I was curious how models perceive speech from people who don\u2019t speak with so-called \u2018standard\u2019 American accents.\u201d<\/p>\n

The team chose accents from five regions that have large populations of English speakers and where the researchers also had connections. Researchers then tested the accents in three different domains \u2014 hiring, presentations and English tests. To study differences at scale, they had AI voices from ElevenLabs<\/a> read the same pieces of text to the AI models being tested. They validated the results with 52 human speakers from the five countries.\u00a0<\/p>\n

After some testing, the team focused the study on three proprietary AI models\u00a0 \u2014 Google\u2019s Gemini 2.5 Pro and Gemini 2.5 Flash, and OpenAI\u2019s GPT-Audio 1.5 \u2014 and one open-source model, Qwen 3 Omni.\u00a0<\/p>\n

\u201cWe wanted to check these state-of-the-art models,\u201d said senior author Yulia Tsvetkov<\/a>, a UW associate professor in the Allen School, \u201cbecause these are what companies would deploy in a real-world setting.\u201d<\/p>\n

The team had the AI voices read various scripts \u2014 interview responses, a class presentation, a conference presentation and English proficiency test responses \u2014 with the five different accents in both male and female voices. The four AI models then rated the responses on a scale of one to seven for both delivery and content, and gave a brief explanation for their scores.\u00a0<\/p>\n

Researchers repeated the interview response test with the 52 human participants, having them read four scripted responses each and respond to questions without a script.\u00a0<\/p>\n

The results favored American and British accents\u00a0\u2014 with a greater disparity between the human accents than between the AI accents.\u00a0<\/p>\n

To better understand the results, researchers analyzed the acoustic features of the accents. They then measured their \u201cacoustic distance\u201d and found that accents further from Americans\u2019 were often ranked lower.\u00a0<\/p>\n

The team found this apparent hierarchy manifested in the models\u2019 explanations, too.\u00a0<\/p>\n

\u201cEven though they repeatedly claimed Chinese and Nigerian speakers \u2018mumbled\u2019 or were \u2018difficult to understand,\u2019 the models were actually very good at transcribing the accented speech,\u201d Ahia said. \u201cThey understood the speakers, but they just had this implicit notion of hierarchy among them based on accent.\u201d<\/p>\n

This bias can seep into models during training in various ways \u2014 possibly through data that describes accented speakers as hard to understand, through data sets that underrepresent certain populations and accents, or through training methods that reinforce conventional expectations about how a professional speaker should sound. English, for instance, is the official language in Nigeria and is the teaching language in Indian schools, yet both accents were rated lower than British and American.<\/p>\n

\u201cOne clear direction for future work is to create more representative training and evaluation data sets that could improve the accuracy and fairness of language models,\u201d Tsvetkov said.\u00a0<\/p>\n

The team plans to expand the research by studying more human accents \u2014 including regional variations within countries \u2014 and by looking at how AI models are analyzing audio in medical contexts, such as automated emergency response.\u00a0<\/p>\n

\u201cFor all of these, we hope that developers build and publish evaluations of their models that don\u2019t focus only on average performance, but also break it down into subpopulations, dialects and so on,\u201d Tsvetkov said. \u201cWe don\u2019t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%.\u201d<\/p>\n

Additional co-authors include Aruna Srivastava<\/a> of Koel Labs, who completed this work as a UW undergraduate in the Allen School; Lucy Li<\/a> of University of Wisconsin\u2013Madison, who completed this research as a UW postdoctoral scholar in the Allen School; Ashley Christendat<\/a> and Farhan Samir<\/a> of the University of Toronto; Sameep Chattopadhyay<\/a>, a UW doctoral student in the Allen School; T\u1eb9j\u00fam\u00e1d\u00e9 \u00c0f\u1ecd\u0300nj\u00e1<\/a> of the CISPA Helmholtz Center; Valentin Hofmann<\/a> of LMU Munich, who completed this research as a UW postdoctoral scholar in the Allen School; Sachin Kumar<\/a> of Ohio State University; and Noah A. Smith<\/a>, UW vice provost for AI and professor in the Allen School.\u00a0<\/p>\n

This research was funded in part by gifts from Google and OpenAI.\u00a0<\/p>\n

For more information, contact Ahia at [email\u00a0protected]<\/a> and Tsvetkov at [email\u00a0protected]<\/a>.<\/p>\n","protected":false},"excerpt":{"rendered":"Newswise \u2014 Artificial intelligence models replicate, and sometimes amplify, all sorts of human biases. In the last few…\n","protected":false},"author":2,"featured_media":957565,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[6810,276,327151,49,48,56437,6808,2727,22867,327152,61,14291],"class_list":["post-957564","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology","tag-all-journal-news","tag-artificial-intelligence","tag-artificial-intelligencebiasaccentslanguage-modelsjob-interviewsspeech-evaluationmultimodal-modelshuman-biasesenglish-proficiency","tag-ca","tag-canada","tag-gender-issues","tag-newswise","tag-public-health","tag-race-and-ethnicity","tag-speech-language","tag-technology","tag-university-of-washington"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/posts\/957564","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/comments?post=957564"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/posts\/957564\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/media\/957565"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/media?parent=957564"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/categories?post=957564"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ca\/wp-json\/wp\/v2\/tags?post=957564"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}