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 state-of-the-art models systematically prefer Standard American English (SAE) continuations even when the preceding context is in AAE, effectively rewriting AAE into SAE. We present an end-to-end framework to audit and mitigate this bias. For auditing, we introduce conditional Dialect Group Invariance (cDGI), which isolate
The increasing sophistication and deployment of Large Language Models (LLMs) are bringing their inherent biases, particularly dialect-based, to the forefront as their societal integration deepens.
This highlights a critical ethical and functional flaw in current state-of-the-art AI, impacting fairness, accessibility, and the practical utility of LLMs for diverse populations.
The systematic 'correction' of dialects like African American English by LLMs will force developers to implement more robust bias auditing and mitigation strategies, moving beyond simple performance metrics.
- · AI ethics researchers
- · Linguistic diversity advocates
- · Open-source AI communities
- · Developers of bias-mitigation tools
- · LLM developers ignoring ethical AI
- · Users of marginalized dialects
- · AI systems lacking inclusive training data
Increased pressure on AI labs to invest in dialect-aware model development and bias detection.
Development of new AI benchmarks and regulatory frameworks specifically targeting linguistic and cultural biases in large models.
Enhanced trust and broader adoption of AI technologies among previously marginalized linguistic groups, leading to new market opportunities and applications.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL