内容来源：https://www.geekwire.com/2026/accent-bias-ai-models-penalize-certain-speakers-in-interviews-and-tests-uw-study-finds/

内容总结：

一项来自美国华盛顿大学的新研究发现，主流AI模型在评估英语口语时会系统性歧视带有口音的人，其中美国口音得分最高，中国和尼日利亚口音得分最低。研究人员测试了谷歌、OpenAI和阿里巴巴的AI模型，让它们对带有美国、英国、中国、印度和尼日利亚五种口音的英语口语进行评分，场景涵盖求职面试、学术会议演讲和英语水平测试。结果显示，模型普遍给美国和英国口音最高分，印度口音居中，中国和尼日利亚口音垫底。这种扣分主要体现在“表达方式”上，但即便所有朗读者读的是同一份文稿、模型被明确告知只评判内容实质，中国和尼日利亚口音的内容得分有时也会偏低。研究共同第一作者、华盛顿大学博士生Orevaoghene Ahia指出，模型反复声称中国和尼日利亚口音者“含糊不清”或“难以理解”，但实际上模型对这些带口音语音的转录能力很好。这一问题的现实影响不容忽视：越来越多企业使用AI筛选面试录像，口语考试也因AI代写问题而日益受到重视，培生PTE学术英语考试等用于签证和移民的英语测试已依赖AI为部分口语作答评分。研究团队强调，他们测试的是企业可能在其基础上构建应用的通用模型，而非招聘平台或考试机构实际使用的产品，后者的运作方式通常不公开，目前也没有已知的AI英语测试公开审计。研究还发现，口音与美式英语的声学差异越大，模型给的分数越低，中国和尼日利亚口音距美式英语最远，印度和英国口音则较接近。研究人员推测，偏见可能来源于训练数据中将带口音者描述为吐字不清、某些口音在数据集中代表性不足，或训练过程强化了“专业人士应如何说话”的刻板印象。值得注意的是，英语是尼日利亚的官方语言，也是印度学校的教学语言，但这两种口音得分仍低于美英口音。该研究于10月25日在布达佩斯举行的自然语言处理经验方法会议上发表，部分经费来自谷歌和OpenAI的资助。研究团队计划下一步研究更多口音及AI在医疗急救等场景中的语音处理表现，并呼吁开发者在发布模型评估时不仅关注平均性能，还应按子群体和方言等维度细分数据。

中文翻译：

设想两名求职者对同一个面试问题给出完全相同的回答，一字不差。一个带美国口音，另一个带尼日利亚口音。让一个领先的AI模型给他们打分，美国人很可能会胜出。

这就是华盛顿大学一项新研究的核心发现。该研究测试了来自谷歌、OpenAI和阿里巴巴的AI模型如何评估带有五种口音的英语：美国、英国、中国、印度和尼日利亚口音。

在求职面试、会议演讲和英语水平测试中，这些模型普遍给美国和英国口音的说话者评分最高，印度口音居中，中国和尼日利亚口音最低。

扣分主要集中在表达方式上。但中国和尼日利亚说话者的内容得分有时也会下降，尽管每位说话者朗读的是同一份稿子，而且模型被明确告知只评判内容实质。

“尽管它们反复声称中国和尼日利亚说话者‘含糊不清’或‘难以理解’，但模型实际上非常擅长转录带口音的语音，”该研究共同第一作者、华盛顿大学保罗·G·艾伦计算机科学与工程学院博士生奥雷瓦奥格内·阿希亚说。

这件事的影响远不止一条差评。企业越来越多地使用AI来筛选录制好的求职面试，而随着书面作业越来越容易被AI伪造，教师们也越来越多地采用口语考试。至少有一项用于签证和移民的主要英语测试——培生PTE学术英语考试——依赖AI为部分口语答案评分。根据论文引用的数据，该考试在2023年约有120万人次参加。

研究人员并未测试招聘平台或考试机构实际使用的产品。这些系统的运作方式通常不公开，论文指出，据作者所知，目前没有对AI评分的英语测试进行过公开审计。相反，团队测试的是企业可以在此基础上构建应用的通用模型。

“我们希望检查这些最先进的模型，因为这些都是企业会在现实环境中部署的，”该研究的资深作者、华盛顿大学艾伦学院副教授尤利娅·茨韦特科夫说。

请听以下五种口音。
美国口音：
尼日利亚口音：
中国口音：
印度口音：
英国口音：

该研究的部分资金来自谷歌和OpenAI的捐赠，它们的模型也在被测试之列。

为了理解这一模式，研究人员测量了每种口音与美国英语在声学上的差异程度。中国和尼日利亚口音差距最大，而模型的评分随着这一差距的增大而下降。印度和英国口音则接近美国基准。

研究人员认为，这种偏见可能来自将带口音的说话者描述为吐字不清的训练数据，来自对口音代表性不足的数据集，或者来自强化“专业人士应该怎样说话”这一观念的训练过程。英语是尼日利亚的官方语言，也是印度学校的教学语言，但两种口音的得分都低于美国和英国口音。

这项研究是华盛顿大学一系列AI偏见研究中的最新一项。此前的项目发现，AI模型偏爱带有白人和男性关联姓名的简历；带有政治倾向的聊天机器人可能改变人们的观点；AI编写的儿童故事几乎抹去了雌性动物角色。

该团队于10月25日在布达佩斯举行的自然语言处理经验方法会议上展示了这项研究。接下来，研究人员计划研究更多口音，包括各国内部的地区变体，并考察AI模型在医疗场景（如自动紧急响应）中如何处理音频。

“对于所有这些，我们希望开发者构建并发布对其模型的评估，不仅关注平均表现，还要按子群体、方言等维度进行细分，”茨韦特科夫说。“我们不希望平均表现对一般人群达到97%的准确率，而对较小群体只有20%。”

英文来源：

Picture two job candidates giving the same answer to the same interview question, word for word. One has an American accent. The other has a Nigerian accent. Ask a leading AI model to score them, and the American will likely come out ahead.
That’s the gist of new research from the University of Washington, which tested how AI models from Google, OpenAI and Alibaba evaluated English spoken with five accents: American, British, Chinese, Indian and Nigerian.
Across job interviews, conference presentations and English proficiency tests, the models generally ranked American and British speakers highest, Indian speakers in the middle, and Chinese and Nigerian speakers lowest.
The penalty fell mostly on delivery. But content scores for Chinese and Nigerian speakers sometimes slipped, too, even though every speaker read the same script and the models were told to judge only the substance.
“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,” said Orevaoghene Ahia, a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering and co-lead author of the study.
The stakes go beyond a bad review. Companies increasingly use AI to screen recorded job interviews, and instructors are turning to oral exams as written work gets easier to fake with AI. At least one major English test used for visas and immigration, Pearson’s PTE Academic, relies on AI to score some spoken answers. It was given about 1.2 million times in 2023, according to figures cited in the paper.
The researchers didn’t test the actual products used by hiring platforms or test makers. How those systems work is usually not public, and the paper notes there are no public audits of AI-scored English tests that the authors know of. Instead, the team tested general-purpose models that companies could build on.
“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 in the Allen School and the study’s senior author.
Listen to the five accents below.
American:
Nigerian:
Chinese:
Indian:
British:
The research was funded in part by gifts from Google and OpenAI, whose models were among those tested.
To understand the pattern, the researchers measured how acoustically different each accent was from American English. Chinese and Nigerian accents were the farthest away, and models’ scores dropped as that distance grew. Indian and British accents sat close to the American baseline.
The researchers suggest the bias may come from training data that describes accented speakers as unclear, from data sets that underrepresent some accents, or from training that reinforces ideas about how a professional should sound. English is an official language in Nigeria and the language of instruction in Indian schools, yet both accents scored below American and British ones.
The study is the latest in a run of UW research on AI bias. Earlier projects found that AI models favored resumes with white- and male-associated names, that politically slanted chatbots could shift people’s views, and that AI-written children’s stories nearly erased female animal characters.
The team presented the research Oct. 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest. Next, the researchers plan to study more accents, including regional variations within countries, and to look at how AI models handle audio in medical settings such as automated emergency response.
“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,” Tsvetkov said. “We don’t want the average performance to be 97% accurate for the general population, but for a smaller group only 20%.”