Linguistic and dialectal bias in large language models: a review of evidence, mechanisms,and implications (20182025)

Large Language Models (LLMs) increasingly mediate communication, evaluation, and decision-making across social and institutional domains. While prior work has documented various forms of bias in language technologies, linguistic variation has received comparatively less integrated treatment despite its deep sociolinguistic significance. This review synthesizes recent research on how LLMs reflect and reinforce linguistic bias toward nonstandard dialects, with a focus on African American English and other marginalized varieties. Drawing on studies spanning natural language understanding, reasoning, speech recognition, and conversational systems, we show that LLMs exhibit consistent performance disparities when processing dialectal inputs, even when semantic content is held constant. Beyond accuracy degradation, emerging evidence demonstrates that dialect functions as a social signal that triggers differential judgment, stereotyping, and covertly racist decision-making in high-stakes contexts. We organize the literature around four themes: dialect-based performance gaps, dialect-triggered social evaluation, measurement frameworks for linguistic bias, and proposed mitigation strategies. While recent benchmarks and adaptation methods improve robustness, we argue that many approaches treat dialect bias as a technical deficiency rather than a sociotechnical problem rooted in language ideology and power. By synthesizing research across technical, social, and evaluative dimensions, this review identifies methodological gaps and proposes directions for developing linguistically and culturally robust LLMs.