Researchers Find AI Models Silently Rewrite African American English
A new study reveals that AI language models often change African American English into Standard American English without users knowing. Researchers have developed a way to detect and reduce this bias in AI systems.
Researchers from arXiv published a study showing that large language models (LLMs) frequently rewrite African American English (AAE) into Standard American English (SAE) without explicit instructions. These models, which power many AI chatbots and writing tools, prefer SAE continuations even when the input is in AAE, effectively altering the dialect. The study tested six different instruction-tuned LLMs (ranging from 14B to 70B parameters) and found this bias across all of them, highlighting a significant issue in how AI interprets and generates language.
This bias matters because it can lead to misunderstandings and erasure of cultural linguistic diversity. For example, if you ask an AI to complete a sentence in AAE, it might automatically switch to SAE, which could be frustrating or even offensive to users who speak AAE. This isn't just about correctness—it's about respecting and preserving different ways of speaking.
The researchers also introduced a new auditing framework called conditional Dialect Group Invariance (cDGI) to detect this bias, and they explored activation steering as a mitigation technique to reduce the silent correction of AAE by LLMs.