How Does "English (US)" Become the Default? Triangulating Structural Bias Towards American English Across the LLM Pipeline
This study investigated how large language models structurally prefer American English (AmE) despite global diversity. By auditing pretraining data, tokenizer designs, and model outputs using triangulation methods, researchers found a systematic bias toward AmE. This bias is deeply embedded in the training data and tokenization systems, leading LLMs to default to AmE even when prompted with British English.