How Does “English (US)”
Become the Default?

Triangulating structural bias towards American English across the LLM pipeline

Mir Tafseer Nayeem1 Davood Rafiei1

1 University of Alberta

Paper Code Dataset
Study overview tracing regional preference through data exposure, representation, and generation, across model origins, prompt conditions, domains, registers, linguistic categories, and alignment confidence.
Study overview Regional preference is examined at three connected stages: what models encounter, how they encode and score variants, and which forms they generate.
Overview

Abstract

How does “English (US)” become a default in large language models? We trace American English (AmE) and British English (BrE) through training data, model representation, and generated language. The study uses 1,813 matched regional variants and introduces DIALIGN, a training-free method for estimating regional alignment from distributional evidence.

Across six pretraining corpora and 21 post-training datasets, AmE variants are more prevalent. AmE forms are generally more compactly tokenized, and BrE counterfactuals receive higher prediction cost under equal-token-count controls. Neutral English prompts produce predominantly AmE outputs; BrE prompting shifts the balance but does not consistently remove the AmE preference.

Contributions

What this study adds

01

A pipeline-wide audit

Regional preference is measured across data exposure, representation, and generated language.

02

A training-free measure

DIALIGN estimates regional alignment from local distributional evidence beyond isolated spelling choices.

03

A matched variant resource

1,813 AmE–BrE pairs support consistent comparisons across the analyses.

Method

A paired comparison across the LLM pipeline

The study builds a curated resource of 1,813 matched American-English and British-English variants. The pairs support comparable measurements across corpora, tokenizers, model likelihoods, and generated responses.

DIALIGN

DIALIGN is a dynamic, training-free method for estimating regional alignment. It aggregates local distributional evidence from overlapping text spans, including lexical, grammatical, structural, stylistic, and multi-word patterns.

On a balanced benchmark of 1,500 news passages, DIALIGN reaches 93.18% accuracy and 93.38 F1.

DIALIGN benchmark evaluation summary: 93.18 percent accuracy, 90.67 precision, 96.25 recall, and 93.38 F1, with confidence-margin distributions.
DIALIGN validation Evaluation on balanced American-English and British-English news passages.

Examples of regional variation

Representative examples from Table 6. These reflect common majority-preference usage and are illustrative, not exhaustive.
LevelAmerican English (AmE)Reference: British English (BrE)
Orthographycolor · center · organize · traveling · favorite · check · jewelry · program · catalog · defensecolour · centre · organise · travelling · favourite · cheque · jewellery · programme · catalogue · defence
Vocabularycell phone · crosswalk · downtown · train station · parking lot · ZIP code · vacation · apartment · elevator · line · sidewalk · flashlight · soda · popsiclemobile phone · zebra crossing · city centre · railway station · car park · postcode · holiday · flat · lift · queue · pavement · torch · fizzy drink · ice lolly
Grammar / usageI just ate · on the weekend · Do you have a pen? · The team is winning · different thanI’ve just eaten · at the weekend · Have you got a pen? · The team are winning · different from
Conventionsdouble quotes · punctuation inside quotes · December 31, 2024 · first floor · 11:15 PM · 68°Fsingle quotes · punctuation outside quotes · 31 December 2024 · ground floor · 11.15 pm · 20°C
Findings

Evidence across data, representation, and generation

The table summarizes the main empirical result at each stage. Regional preference is measured through different signals at each stage, so the values should be read within their respective analyses.

Summary of the paper’s main findings
StageCoverageKey result
Data exposure6 pretraining corpora; 21 post-training datasetsAll favor AmE. AmE accounts for 76.17% of 28.86 million matched-variant occurrences in the post-training data.
Representation9 tokenizers; 10 model checkpointsAmE is generally more compactly tokenized. BrE counterfactuals receive higher per-token loss in all 10 checkpoints under equal-token-count controls.
Generation10 models; Natural Questions and ELI5AmE is the majority under neutral English prompts. British-English prompting moves outputs toward BrE, with effects that vary by model and register.

The matched-pair design supports a consistent cross-stage comparison. These findings establish a recurring association; they do not isolate the causal contribution of each stage.

Implications

Defaults are built into systems

The results point to a pipeline-wide pattern rather than a generation-only quirk. Data composition, tokenizer design and lineage, post-training choices, and evaluation practices can all shape which variety is treated as ordinary.

This evidence motivates targeted audits and interventions at each stage, and raises questions about linguistic homogenization, representation, and equitable deployment. The study establishes consistent associations across the evaluated pipeline; it does not claim that any single stage alone causes the downstream preference.

Paper

Paper

How Does “English (US)” Become the Default? Triangulating Structural Bias Towards American English Across the LLM Pipeline.

@misc{nayeem2026englishdefault,
  title = {How Does ``English (US)'' Become the Default?},
  author = {Nayeem, Mir Tafseer and Rafiei, Davood},
  year = {2026}
}