A University of Washington-led team of researchers found that state-of-the-art AI models were biased against speakers with certain accents when evaluating human speech in job interviews, conference presentations and spoken English proficiency tests. The team investigated multimodal language models, including two from Google and one from OpenAI, scoring responses from voices speaking English with American, British, Chinese, Indian and Nigerian accents. The models generally ranked Chinese and Nigerian accents the lowest, Indian in the middle, and American and British the highest; even when content was identical and the models could accurately transcribe the speech, the systems ranked both delivery and content higher for American and British speakers. The team presented its research Oct. 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest, Hungary. Co-lead author Orevaoghene Ahia, a UW doctoral student, noted people are trusting audio AI models as third-party evaluators in hiring, oral exams and certain English tests in immigration, contexts where fairness matters. Researchers tested the accents in hiring, presentations and English tests, had AI voices from ElevenLabs read the same texts, and validated results with 52 human speakers from the five countries. The study focused on Google’s Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI’s GPT-Audio 1.5 and the open-source Qwen 3 Omni.