PulseAugur
EN
LIVE 11:34:10

LLMs rewrite African American English to Standard American English, new study finds

By PulseAugur Editorial · [2 sources] ·

A new research paper details how large language models (LLMs) systematically alter African American English (AAE) into Standard American English (SAE), effectively rewriting the dialect. The study introduces a framework for auditing this bias using conditional Dialect Group Invariance (cDGI) and identifies negative concord as a key trigger. For mitigation, the researchers applied activation steering, a training-free method, which significantly reduced bias while maintaining SAE fluency. The work also includes the release of REAL-AAE, a substantial parallel corpus of AAE and SAE text. AI

IMPACT Highlights a significant bias in LLMs that could impact communication and representation for millions of speakers.

RANK_REASON The cluster contains a research paper detailing findings on LLM bias and a proposed mitigation method.

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

LLMs rewrite African American English to Standard American English, new study finds

COVERAGE [2]

  1. arXiv cs.CL TIER_1 English(EN) · Huan Wu, Ali Emami, Muhammad Furquan Hassan, Osaretin Igbinoba, Osakpolor Idusuyi, Osamede Igbinoba, Faiza Khan Khattak, Laleh Seyyed-Kalantari ·

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

    arXiv:2607.06845v1 Announce Type: new Abstract: African American English (AAE), a rule-governed dialect spoken by over 30 million people, is routinely misinterpreted and "corrected" by large language models (LLMs). Across six instruction-tuned LLMs (14B to 70B), we show that stat…

  2. arXiv cs.CL TIER_1 English(EN) · Laleh Seyyed-Kalantari ·

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

    African American English (AAE), a rule-governed dialect spoken by over 30 million people, is routinely misinterpreted and "corrected" by large language models (LLMs). Across six instruction-tuned LLMs (14B to 70B), we show that state-of-the-art models systematically prefer Standa…

RELATED ENTITIES

RELATED TOPICS