LLMs Silently Correct African American English: Auditing and Mitigating Dialect Bias via Activation Steering

Huan Wu and colleagues found that six instruction-tuned LLMs systematically rewrite African American English into Standard American English. The researchers introduced an auditing framework that identifies syntactic constructions like negative concord as universal bias triggers across all tested models. They propose activation steering, a training-free method that reduces dialect bias 5 to 20 times more effectively than prompting while preserving SAE fluency. They also released REAL-AAE, the largest real-AAE parallel corpus to date, comprising 17,479 validated triplets.