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

Short summary

Six instruction-tuned LLMs (14B–70B) systematically rewrite African American English into Standard American English, with syntactic constructions like negative concord acting as universal bias triggers. The authors introduce conditional Dialect Group Invariance for auditing and activation steering for mitigation, reducing bias 5–20x more than prompting while preserving SAE fluency. They also release REAL-AAE, the largest real-AAE parallel corpus with 17,479 triplets validated by native speakers.

• •LLMs systematically rewrite AAE into SAE across 6 models (14B–70B)
• •Activation steering reduces dialect bias 5–20x more than prompting, training-free
• •REAL-AAE corpus released: 17,479 AAE/SAE triplets, largest of its kind

Generated with AI, which can make mistakes.