By skyforbes 
Picture this: a young teacher from rural Alabama logs into ChatGPT to polish her resume for a school district interview. She’s excited, typing out her experience in her natural dialect, only to get back responses laced with condescension, like the AI is schooling her on “proper” English [2][3]. What she doesn’t realize is that this isn’t just a glitch, it’s a deep-seated bias baked into the model, turning a helpful tool into an unwitting gatekeeper of discrimination [6][7].
Researchers at UC Berkeley dove into this nightmare, leading a charge to expose how ChatGPT treats different English dialects [8]. Their work uncovers a harsh reality: AI isn’t the neutral savior we hoped for, it’s amplifying real-world prejudices in subtle, sneaky ways [5].
The Spark That Ignited the Investigation
Eve Fleisig and her team at the Berkeley Artificial Intelligence Research lab started with a simple hunch. They noticed how language models, trained mostly on “standard” American and British English, fumble when faced with dialects like African American English or Southern American English [2]. This wasn’t abstract theory, it was personal, drawing from Fleisig’s own background in linguistics and a drive to make AI fairer for everyone [6].