“Accent bias: AI models penalize certain speakers in interviews and tests, UW study finds”

Key points

• UW researchers tested AI models from Google, OpenAI and Alibaba on five accents.
• American and British accents ranked highest, Indian in the middle, Chinese and Nigerian lowest.
• Bias mainly affected delivery scores, though content scores sometimes dropped for Chinese and Nigerian speakers.
• Study emphasizes that many AI scoring systems are not audited and may reflect biases.
• Researchers plan to expand tests to more accents and medical audio contexts.