PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Mir Tafseer Nayeem, Davood Rafiei

arXiv:2604.04204v1cs.CLcs.AIcs.CYcs.ETcs.LG

TL;DR

LLMs expose limited language settings, and it remains unclear which English varieties they implicitly prefer. This paper compares AmE and BrE using a curated variant corpus, DIALIGN, and audits across pretraining data, tokenizers, and model outputs. Across these stages, the evidence identifies AmE as the default and BrE as consistently disadvantaged.

  • Problem

    The paper investigates whether LLMs reproduce asymmetries between AmE and BrE, an underexplored question given limited language settings and the models' wide deployment.

    Source: §1 · pp. 1–2

  • Method

    The study combines a curated AmE–BrE variant corpus and DIALIGN with audits of pretraining corpora, tokenizer representations, and generative outputs.

    Source: §1 · p. 3, §3.1 · p. 4

  • Results

    AmE emerges as the default while BrE is consistently disadvantaged, with skew reported in pretraining corpora, tokenizer encoding, and model generations.

    Source: §4 · p. 5, §5 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    The findings raise concerns about linguistic homogenization and epistemic injustice and motivate more balanced corpora, dialect-sensitive tokenizers, and alignment strategies.

    Source: §9 · p. 10

    The analysis is English-specific, and extensions to other languages must adapt its segmentation assumptions to local linguistic norms.

    Source: Ethics Statement · p. 10

Abstract

from arXiv · show

Large language models (LLMs) are increasingly deployed in high-stakes domains, yet they expose only limited language settings, most notably "English (US)," despite the global diversity and colonial history of English. Through a postcolonial framing to explain the broader significance, we investigate how geopolitical histories of data curation, digital dominance, and linguistic standardization shape the LLM development pipeline. Focusing on two dominant standard varieties, American English (AmE) and British English (BrE), we construct a curated corpus of 1,813 AmE--BrE variants and introduce DiAlign, a dynamic, training-free method for estimating dialectal alignment using distributional evidence. We operationalize structural bias by triangulating evidence across three stages: (i) audits of six major pretraining corpora reveal systematic skew toward AmE, (ii) tokenizer analyses show that BrE forms incur higher segmentation costs, and (iii) generative evaluations show a persistent AmE preference in model outputs. To our knowledge, this is the first systematic and multi-faceted examination of dialectal asymmetries in standard English varieties across the phases of LLM development. We find that contemporary LLMs privilege AmE as the de facto norm, raising concerns about linguistic homogenization, epistemic injustice, and inequity in global AI deployment, while motivating practical steps toward more dialectally inclusive language technologies.

1 Introduction

The paper asks which English variety LLMs implicitly prefer, focusing on American and British English. It frames the comparison as structural bias and introduces a multi-stage audit of the LLM development pipeline.

1 Introduction

  • The study asks which English variety LLMs implicitly prefer, focusing on AmE and BrE as influential, well-documented standards for controlled analysis.Source: §1 · p. 1
  • BrE retains normative prestige and institutional standing in many regions, while AmE dominates global culture and digital communication as a de facto norm.Source: §1 · p. 2
  • The paper frames potential preference as structural bias and traces asymmetries across pretraining corpora, tokenizers, and generative behavior rather than only downstream task disparities.Source: §1 · p. 2
  • The authors construct a curated corpus of 1,813 parallel AmE–BrE lexical variants to support consistent, high-precision analysis.Source: §1 · p. 2
  • DIALIGN estimates dialectal alignment from distributional evidence, and the study triangulates evidence across pretraining corpora, tokenizer representations, and generative preferences.Source: §1 · p. 3

2 Postcolonial Framing and Scope

The paper uses postcolonial theory as an interpretive lens for studying AmE–BrE asymmetries. Its scope is limited to these varieties because the available reference corpus distinguishes them, while other English varieties remain relevant to the broader concern.

2 Postcolonial Framing and Scope

  • The postcolonial lens is an interpretive framework that motivates the study's scope, justifies the AmE–BrE comparison, and explains the significance of observed asymmetries.Source: §2 · p. 3
  • BrE became embedded in governance, education, and law across many former colonies and often persisted after independence, providing historical context for the comparison.Source: §2 · p. 4
  • The comparison is operationally necessary because Google Books Ngram distinguishes only AmE and BrE; the authors note implications for other varieties, including Indian, Australian, New Zealand, and Canadian English.The paper presents this work as a pairwise comparison, not an analysis of all English varieties.Source: §2 · p. 4

3 Methodology: Resources and Alignment

The methodology pairs a curated AmE–BrE variant resource with DIALIGN, a training-free method for estimating a text's dialectal alignment. DIALIGN uses distributional evidence to produce alignment probabilities.

3.1 DIALIGN: Dialectal Alignment Score

  • DIALIGN is a dynamic, training-free method that estimates a text's AmE or BrE alignment from distributional evidence across lexical, grammatical, structural, stylistic, and multi-word contrasts.Source: §3 · p. 4, §3.1 · p. 4

3 Methodology: Resources and Alignment

  • The authors compile 1,813 parallel AmE–BrE lexical variants from linguistic sources and web-based lexicons, covering spelling and vocabulary differences.Source: §3 · p. 4

3.1 DIALIGN: Dialectal Alignment Score

  • DIALIGN extracts 2–5-grams, compares their Google Books Ngram frequencies, signs each log-ratio by dialect preference, aggregates weighted evidence, and normalizes scores into alignment probabilities.Source: §3.1 · pp. 4–5

4 RQ1: Auditing Dialectal Skew in Pretraining Corpora

The audit finds statistically significant AmE skew across six pretraining corpora, with stronger imbalance in spelling than vocabulary variants. DIALIGN also indicates broader AmE preferences.

4 RQ1: Auditing Dialectal Skew in Pretraining Corpora

  • The audit combines variant-pair token distributions with DIALIGN to assess lexical, structural, stylistic, grammatical, and multi-word contrasts.Pairwise frequency differences are tested with Wilcoxon signed-rank tests; all corpora have p-values below 0.01.Source: §4 · p. 5
  • The authors interpret consistent asymmetry across these measures as dialectal skew and evidence of structural bias in pretraining corpora.Table 2 reports corpus-level lexical distributions and DIALIGN scores, with bracketed mean confidence scores.Source: §4 · p. 5, Table 2 · p. 6
  • Across six pretraining corpora, all show significant AmE skew; spelling margins exceed 70%, while vocabulary differences are less extreme but consistently favor AmE.DIALIGN likewise indicates a clear AmE preference, often with high confidence.Source: §4 · p. 5, Table 2 · p. 6
  • Figure 2 shows stronger AmE clustering for orthographic variants than vocabulary variants, which remain AmE-leaning.The figure summarizes variant probabilities across three corpora; additional corpora appear in Appendix Figure 4.Source: Figure 2 · p. 6, §4 · p. 6

5 RQ2: Quantifying Representation in Regional Tokenizers

The tokenizer analysis tests whether AmE and BrE variants receive equally efficient representations. Across the analyzed tokenizers, BrE forms have higher fertility and more frequent heavy segmentation, especially for vocabulary differences.

5 RQ2: Quantifying Representation in Regional Tokenizers

  • The research question is whether publicly available regional tokenizers encode AmE and BrE variants with equal efficiency.The paper notes that tokenizer design may introduce dialectal skew before model inference or generation.Source: §5 · p. 6
  • The study measures fertility, the average subword tokens per word, and granularity, the distribution of words split into 1, 2, 3, or more subwords.Fertility gives a mean-level view; granularity captures long-tail fragmentation.Source: §5 · pp. 6–7
  • Across tokenizers, BrE fertility is higher; the vocabulary gap is ∆v = 18.72%, versus orthographic gaps of ∆o = 2.85% to 5.42%.Table 3 reports fertility by tokenizer and variant category; lower fertility indicates more efficient tokenization.Source: Table 3 · p. 7, §5 · p. 7
  • BrE variants appear more often in the 3+ subword bin, while Falcon and Gemma tokenize more compactly and StableLM mirrors GPT-4, likely through tokenizer reuse.Figure 3 compares counts of variant pairs split into 1, 2, or 3+ subwords.Source: Figure 3 · p. 7, §5 · p. 8
  • European and Chinese tokenizers show better BrE coverage; Velvet alone favors BrE orthography (∆o = −0.69%), and DeepSeek has the lowest vocabulary skew (∆v = 12.66%).Gemma has the lowest overall fertility across both dialects, with a 262K vocabulary; the text suggests controlled expansion may improve balance.Source: §5 · pp. 7–8

6 RQ3: Evaluating Dialectal Preferences in LLM Generation

The evaluation finds AmE to be the dominant generative default across models, persisting even when prompts explicitly request British English. BrE uptake is greater in informal ELI5 than in formal Natural Questions.

Experimental Setup

  • The evaluation estimates whether each generated response aligns more closely with AmE or BrE using DIALIGN.Given a question and a model-generated response, DIALIGN estimates dialectal alignment.Source: §6 · p. 8
  • Questions containing AmE or BrE variants were removed to avoid lexical priming, and each question elicited generations under English and British English (en-GB) conditions.The evaluation covers formal Natural Questions and informal ELI5; results report AmE-output percentages and mean AmE confidence.Source: §6 · p. 8, Table 4 · p. 8

Results & Analysis

  • Under default English prompts, most models produce 65–80% AmE outputs, and AmE rarely falls below 40% under British English prompts.The study reports both AmE-classified output percentages and mean AmE alignment confidence; Table 4 covers Natural Questions and ELI5.Source: §6 · p. 8, Table 4 · p. 8
  • ELI5 shows greater BrE uptake than Natural Questions; LLaMA-3 drops to ∼30% AmE under en-GB, while formal contexts reinforce AmE norms.The authors interpret conversational registers as providing more lexical flexibility than formal contexts.Source: §6 · p. 8

Key Takeaways

  • AmE remains the entrenched generative default even under BrE prompts, while BrE uptake is stronger in informal domains than formal ones.The authors identify pretraining data and tokenizer design as jointly shaping dialectal bias.Source: §6 · p. 8

7 Discussion & Recommendations

The discussion connects dialectal skew to broader cultural and epistemic asymmetries, then recommends changes to pretraining data and tokenizer design.

Dialectal Skew and Broader Implications

  • Models that privilege AmE may carry forward its norms, values, and worldviews, shaping which knowledge is legitimized and which practices are marginalized.The authors connect this concern to hegemonic perspectives encoded in training corpora and to AI governance and Sovereign AI initiatives.Source: §7 · p. 9

Balancing Pretraining Data for Improved Dialectal Representation

  • Pretraining corpora can be enriched by selectively sampling .uk domains, while .ca and .in sources can broaden coverage of World Englishes built on BrE.When naturally occurring data are scarce, synthetic data may be considered, but the authors warn it risks defaulting to AmE; DIALIGN can check synthetic samples for BrE alignment.Source: §7 · p. 9

Dialect-Sensitive Tokenizer Design

  • Dialect-sensitive vocabulary expansion can add BrE tokens absent from base tokenizers, using the AmE–BrE lexicon and granularity-based diagnostics to identify them.The authors identify tokenizer reuse without addressing dialectal asymmetries as another source of bias.Source: §7 · p. 9

8 Related Work

Related work covers data auditing and curation, tokenizer fairness, and dialect robustness, while identifying standard AmE–BrE asymmetries across the LLM pipeline as underexplored.

Pretraining data audits and curation

  • Prior corpus research examines dataset quality and provenance, construction recipes, register-aware sampling, and reuse of filtered web text; this work adds AmE–BrE variation as representational skew.The authors trace this dimension from pretraining corpora into tokenization and generation.Source: §8 · p. 9

Tokenizer fairness

  • Tokenizer-fairness studies report uneven segmentation of equivalent strings and links to efficiency, context budget, cost, and social bias; this work extends the inquiry to intra-English variation.The paper reports that tokenizers encode AmE and BrE forms unevenly.Source: §8 · p. 9

Dialect robustness in NLP tasks

  • Prior dialect-robustness research documents NLP performance gaps and negative stereotypes for AAE and SAsE, while standard varieties such as AmE and BrE remain underexplored.The paper addresses that gap by tracing AmE–BrE asymmetries across the LLM development pipeline through a postcolonial framing.Source: §8 · pp. 9–10

9 Conclusion

The paper concludes that AmE is the default across pretraining corpora, tokenizers, and model outputs, while BrE is consistently disadvantaged.

9 Conclusion

  • Triangulating pretraining-corpus audits, tokenizer representations, and generative outputs, the paper finds AmE emerges as the default while BrE is consistently disadvantaged.The authors frame the findings as raising risks of linguistic homogenization and epistemic injustice, and motivating more balanced corpora, dialect-sensitive tokenizers, and alignment strategies.Source: §9 · p. 10

Ethics Statement

The ethics statement frames the work as an audit of structural linguistic bias with a constructive aim: supporting more inclusive, transparent, and dialect-aware language technologies. It limits the study to public data and excludes named entities from DIALIGN.

Ethics Statement

  • The study uses publicly available corpora and a publicly sourced AmE–BrE lexicon intended for non-commercial research release under CC BY-NC-SA 4.0.It involves no human subjects, personal data, or sensitive attributes.Source: Ethics Statement · p. 10
  • The analysis is an English-specific instantiation; applying it to other languages requires adapting segmentation assumptions to local linguistic norms.The authors note that some languages lack clear orthographic word boundaries or segment linguistic units differently.Source: Ethics Statement · p. 10
  • The work aims to document structural bias and support more inclusive language technologies, while excluding named entities from DIALIGN to reduce privacy or reputational risks.The authors identify risks of epistemic injustice and linguistic homogenization in global AI deployment.Source: Ethics Statement · p. 10

Supplementary Material: Appendices

The appendices situate AmE and BrE in distinct histories of standardization and describe the lexicon and DIALIGN procedures used to compare them. They also illustrate how the method uses distributional evidence to measure dialectal alignment.

A Brief History

  • AmE and BrE standards emerged through distinct codification histories, while British colonial institutions and later U.S. cultural and economic influence gave each variety different forms of global reach.The appendix traces British spelling codification to Johnson’s 1755 dictionary and American spelling advocacy to Webster’s 1828 dictionary.Source: A Brief History · p. 20

B Dialectal Variant Corpus: Details

  • The curated reference set contains 1,813 parallel AmE–BrE word pairs, retaining strict one-to-one mappings for consistent word-level comparisons.It covers spelling and vocabulary differences; multiword and other non-one-to-one mappings are excluded.Source: B Dialectal Variant Corpus: Details · p. 20

B.1 Dialectal Variant Grouping

  • A deterministic procedure assigns each pair to one of ten mutually exclusive groups, then maps those groups to spelling, vocabulary, or uncategorized classes.The grouping scheme is described as exhaustive and provides the basis for analyzing AmE–BrE variants.Source: §B.1 · p. 21

C Implementation Details of DiAlign

  • DIALIGN uses n-gram frequency evidence and lexicon-based boosting to estimate alignment; its parameter choices are a reproducible baseline, not a uniquely optimal configuration.The illustrated scoring uses Google Books Ngram frequencies for AmE and BrE and returns alignment probabilities; alternative parameter settings may yield different or improved performance.Source: C Implementation Details of DiAlign · pp. 22–23, D Walkthrough of DiAlign with… · p. 24

D Walkthrough of DiAlign with Illustrative Input

  • The walkthrough combines orthographic, lexical, grammatical, and multiword contrasts so DIALIGN can capture structural and idiomatic patterns beyond single-word substitutions.It segments parallel AmE and BrE texts into contiguous 2–5-grams, queries Google Books Ngrams, and normalizes the evidence into alignment probabilities.Source: D Walkthrough of DiAlign with… · pp. 24–25, Figure 8 · p. 24

E Details of the Meta-Evaluation of DiAlign

DIALIGN is meta-evaluated on balanced samples from U.S. and U.K. news sources, and the appendices document the generation setup and corpus-audit procedures. It achieves strong classification performance, while the source-based benchmark may also capture journalistic differences.

E Details of the Meta-Evaluation of DiAlign

  • The meta-evaluation uses 750 BBC-derived BrE summaries and 750 HuffPost-derived AmE samples, with no standard dataset providing explicit AmE–BrE annotations.The source provenance supplies dialectal signals, but differences in journalistic style may add variation beyond dialect.Source: E Details of the Meta-Evaluation of… · pp. 25–27
  • On the balanced AmE–BrE benchmark, DIALIGN achieves 93.2% accuracy, 90.7% precision, 96.3% recall, and 93.4 F1, with complementary gains from divergence weighting and boosting.Correct predictions are mostly high-confidence, while errors cluster near the decision boundary. Removing both components causes the largest ablation drop.Source: E Details of the Meta-Evaluation of… · p. 26, Figure 10 · p. 26, Table 6 · p. 27

F Details of the Experimental Setup for RQ3

  • The generation evaluation compares outputs under English and British English (en-GB) prompts across formal Natural Questions and informal ELI5, scoring dialectal alignment rather than factual correctness.Questions containing dialect variants are removed to reduce lexical priming, and 600 questions are sampled across the two datasets.Source: F Details of the Experimental Setup… · pp. 27–28

G Details of Pretraining Datasets

  • Because querying Google Books Ngrams for web-scale corpora is computationally prohibitive, the audit samples 10,000 documents from each of six pretraining datasets.The described corpora include books, encyclopedic text, web data, and mixed-domain collections.Source: G Details of Pretraining Datasets · pp. 28–29

H Analysis of Word-Length Adherence

  • Closed-source models such as GPT-4o and Gemini stay near the 50-word target, with standard deviations around 2 and ranges around 45–60 words.The appendix reports that adherence varies across models despite a shared length instruction.Source: H Analysis of Word-Length Adherence · p. 29, Table 7 · p. 29

I Preprocessing Pipeline for Auditing Pretraining Corpora

  • The corpus-audit preprocessing standardizes text by lowercasing, removing markup and non-alphabetic characters, and normalizing whitespace before counting lexicon variants.These steps are intended to ensure consistent variant counting across corpora.Source: I Preprocessing Pipeline for Auditing… · p. 30

J Limitations & Future Directions

The section bounds the study's dialect focus, lexicon, evaluation settings, and DIALIGN, while situating its AmE–BrE analysis within related work on corpus audits, tokenization, and dialect robustness. It also discloses the limited assistive uses of LLMs in the research process.

J Limitations & Future Directions

  • The analysis covers AmE and BrE, not other World Englishes or multilingual settings, and its findings may be a lower bound for postcolonial and local non-standard varieties.The authors describe the comparison as a controlled, high-precision foundation for future extensions.Source: J Limitations & Future Directions · p. 30
  • The lexicon's strict one-to-one pairs reduce breadth, and DIALIGN cannot capture asymmetries beyond the AmE–BrE contrast or its reference data.The static lexicon may miss emerging or domain-specific terms; DIALIGN broadens coverage to common structural, stylistic, grammatical, and multi-word contrasts.Source: J Limitations & Future Directions · pp. 30–31
  • Generation tests cover open-domain QA with 50-word responses in formal and informal registers, not dialogue, long-form, or domain-specific contexts.Filtering queries with explicit dialect markers avoids priming, but leaves dialect-cued inputs unexplored; code-switching and creative writing may produce different outcomes.Source: J Limitations & Future Directions · p. 31
  • DIALIGN may return neutral or undefined scores for generic passages, miss subtler stylistic cues, and reflect written-usage and news-style confounds.Its meta-evaluation used BBC and HuffPost news as proxies, and its reliance on Google Books n-grams may underrepresent contemporary internet discourse or emerging slang.Source: J Limitations & Future Directions · p. 31

K Extended Related Work

  • Prior corpus-audit and curation work motivates treating dialect as a corpus-level representation dimension; this study quantifies AmE–BrE distributions and links imbalance to tokenization and generation.The section also frames representational balance alongside quality, domain coverage, register, provenance, duplication, and licensing.Source: K Extended Related Work · p. 31
  • Prior tokenizer research documents disparities and mitigation strategies, while this study extends the inquiry to AmE–BrE differences in segmentation length and consistency.The section discusses interventions including token-embedding fine-tuning and parity-aware byte-pair encoding.Source: K Extended Related Work · p. 32
  • Dialect robustness work has emphasized marginalized or low-resource varieties, especially AAE and South Asian Englishes; this study extends the fairness inquiry to standard AmE and BrE.The authors interpret the pipeline analysis through a postcolonial framing of geopolitical histories, digital dominance, and linguistic standardization.Source: K Extended Related Work · pp. 32–33

L Usage of Large Language Models

  • LLMs assisted with author-reviewed text polishing and code scaffolds; outputs were manually verified and tested, with authors responsible for the content and conclusions.The declared uses were limited to polishing grammar, spelling, phrasing, and word choice, and generating small preprocessing and filtering snippets.Source: L Usage of Large Language Models · p. 33

Related papers

Loading 2604.04204v1…