University of Washington study finds AI models penalize Chinese and Nigerian accents

The study, led by doctoral student Orevaoghene Ahia and associate professor Yulia Tsvetkov, notes that the models transcribed the accented speech accurately but still judged it as “mumbled” or “difficult to understand.” Funding came in part from gifts from Google and OpenAI. The researchers point to training data that under‑represents certain accents or describes them as unclear as possible sources of bias. They stress the stakes for companies that use AI‑based interview screening and for English tests such as Pearson’s PTE Academic, which was administered about 1.2 million times in 2023.

The team plans to extend the work to more regional accents and to explore AI handling of audio in medical settings, urging developers to publish evaluations that break performance down by dialect and subpopulation.