Which English Do LLMs Prefer? Triangulating Structural Bias Towards American English in Foundation Models
This paper investigates and demonstrates through a multi-stage analysis of pretraining corpora, tokenizers, and generative outputs that large language models exhibit a systematic structural bias favoring American English over British English, thereby reinforcing linguistic homogenization and epistemic injustice in global AI deployment.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Picture: The "American Default" Problem
Imagine you walk into a global library that claims to speak every language. You ask the librarian, "Can you speak British English?" The librarian smiles, nods, and says, "Of course!" But then, they hand you a book written entirely in American English, with American spellings, American idioms, and American cultural references. They insist, "This is just English."
This is exactly what this paper argues is happening with Large Language Models (LLMs) like ChatGPT, Claude, and Llama. Even though these models are supposed to be global, they have a hidden, structural bias that treats American English (AmE) as the "standard" and British English (BrE) as a slightly broken or expensive version of it.
The authors, Mir Tafseer Nayeem and Davood Rafiei, decided to investigate why this happens by looking at the entire "factory" that builds these AI models. They call this structural bias.
The Investigation: A Three-Stage Factory Tour
To find the root cause, the researchers didn't just look at the final product (the chatbot's answers). They toured the entire factory, looking at three specific stages where the bias gets baked in.
Stage 1: The Raw Ingredients (Pretraining Corpora)
The Analogy: Imagine baking a cake. If you buy a bag of flour that is 80% white flour and only 20% whole wheat, your cake will naturally taste like white flour, no matter how good the baker is.
The Finding: The researchers audited six massive datasets (the "flour bags") used to train AI. They found that these datasets are heavily skewed toward American content.
- The Result: In these data oceans, American spellings (like color and organize) appear far more often than British ones (colour and organise).
- The Metaphor: The AI is being fed a diet of American media, websites, and books. It's like growing up in a house where only American TV is on; naturally, you start speaking like that.
Stage 2: The Chopping Block (Tokenizers)
The Analogy: Imagine you are a chef chopping vegetables. If your knife is designed to chop carrots easily but struggles with parsnips, you will end up chopping the parsnips into tiny, inefficient pieces, while the carrots come out in perfect chunks. This makes the parsnips harder to cook and takes up more space in your pan.
The Finding: Before an AI can understand a word, it breaks it down into smaller chunks called "tokens." The researchers found that the "knives" (tokenizers) used by these models chop British words into more pieces than American words.
- Example: The word colour might be chopped into three tokens, while color is just one or two.
- The Consequence: Because British words are "chopped" more, they cost the AI more "brain power" (computational tokens) to process. This makes the AI slightly slower and less efficient when using British English, subtly discouraging it from using those forms.
Stage 3: The Final Dish (Generative Outputs)
The Analogy: You finally sit down to eat the cake. Even if you asked the chef to make a "British-style" cake, the result still tastes like the American one because the ingredients were American and the knife struggled with British spices.
The Finding: When the researchers asked the AI to write in British English, the models often slipped back into American English.
- The Stats: Even when explicitly told to use British English, models like GPT-4 and Llama still produced American spellings and vocabulary about 40% to 70% of the time.
- The Metaphor: It's like asking a person who grew up in Texas to speak with a Scottish accent. They might try, but they will inevitably slip back into their native Texas drawl because that's what their brain is wired to do.
Why Does This Matter? (The "Postcolonial" Lens)
The paper uses a historical perspective called Postcolonialism to explain why this bias exists.
- The History: British English was spread by the British Empire (colonization), making it the standard in places like India, Nigeria, and Singapore.
- The Shift: In the 20th century, American culture (Hollywood, the internet, tech giants) took over the global stage.
- The Problem: The AI models are built on the internet, which is currently dominated by American culture. By treating American English as the "default," the AI is essentially saying, "Your way of speaking is the 'real' English, and your way is just a variation."
This creates Epistemic Injustice. It means that millions of people who speak British English (or Englishes influenced by it, like Australian or Indian English) are being told their language is "less efficient" or "less correct" by the machines they rely on for work, school, and law.
The Solution: What Can We Do?
The authors suggest three practical fixes to make AI more fair:
- Better Ingredients: When building the training data, we need to actively seek out and include more British and Commonwealth content, not just American websites.
- Sharper Knives: We need to redesign the "tokenizers" (the chopping tools) so they don't penalize British words. We should give British words the same "chunk size" as American ones.
- New Tools: The authors created a tool called DIALIGN. Think of this as a "dialect detector." It helps developers check if their AI is accidentally favoring American English so they can fix it before releasing the model.
The Takeaway
This paper isn't just about spelling differences like color vs. colour. It's about power. It shows that the technology we use every day is quietly reinforcing a specific cultural viewpoint (American) as the "normal" one, while making other valid ways of speaking feel like errors. By fixing the data, the tools, and the design, we can build AI that truly respects the diversity of the English-speaking world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.