Which English Do LLMs Prefer? Triangulating Structural Bias Towards American English in Foundation Models

A new research study reveals that major large language models exhibit systematic bias toward American English over British English across training data, tokenization, and outputs. The research introduces DiAlign, a method for measuring dialectal alignment, and finds evidence of linguistic homogenization that could impact global AI equity.

• →Six major pretraining corpora show systematic skew toward American English over British English varieties.
• →LLM tokenizers impose higher segmentation costs on British English forms compared to American English.
• →Generative AI models consistently prefer American English in their outputs despite global English diversity.
• →The study introduces DiAlign, a training-free method for measuring dialectal bias in language models.
• →Researchers warn of linguistic homogenization and epistemic injustice in global AI deployment.