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

Mir Tafseer Nayeem; Davood Rafiei

cs.CLApr 5, 2026

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

Authors: Mir Tafseer Nayeem, Davood Rafiei

Organizations: Department of Computing Science University of Alberta

Abstract

Large language models (LLMs) are increasingly embedded in educational, professional, and public infrastructure, yet widely used platforms expose "English (US)" as a primary English setting despite the global diversity of English. We ask: How does "English (US)" become the default? We study this question as structural bias, examining how geopolitical histories of data curation, digital dominance, and linguistic standardization intersect with the LLM development pipeline. Using British English as a controlled reference, we construct a curated resource of 1,813 matched American English (AmE)--British English (BrE) variants and introduce DiAlign, a dynamic, training-free method for estimating regional alignment from distributional evidence. We triangulate the AmE preference across data exposure --> representation --> generation, jointly examining pretraining and post-training data, tokenizer behavior and provenance, model prediction cost, and generated language across developer countries, prompt conditions, domains and sources, linguistic categories, and registers. AmE is consistently favored across all six audited pretraining corpora and 21 post-training datasets, is generally represented more compactly by tokenizers, and receives lower prediction cost. It also remains the dominant generation default under neutral English prompting; British-English prompting shifts this preference toward BrE but does not consistently eliminate the AmE default. To our knowledge, this is the first rigorous pipeline-wide study of structural bias across major phases of LLM development. Our findings show that contemporary LLMs privilege AmE as the de facto norm, raising concerns about linguistic homogenization, epistemic injustice, and inequity in global AI deployment, while providing a rigorous basis for targeted component-level intervention.

Figures & tables

Figure 1: Triangulating how “English (US)” becomes the default across the LLM pipeline. We trace this asymmetry across three connected stages: data exposure $\rightarrow$ representation $\rightarrow$ generation , using British English as the controlled reference, Ref(BrE). The study jointly examines pretraining and post-training data, tokenizer behavior and provenance, model prediction cost, and generated language. This triangulation reveals where the American English (AmE) advantage first becomes visible, whether it persists across components, and how it manifests in model outputs. Its strength and form are examined across pipeline stages, developer countries (e.g., USA, UK, China, France, Italy, and UAE), prompt conditions (e.g., default English vs. British English), domains and sources (e.g., web, books, encyclopedic text, and mixed-domain corpora), linguistic categories (e.g., orthographic vs. vocabulary contrasts), registers (e.g., formal/encyclopedic vs. informal/conversational), and alignment confidence levels. This breadth provides a deeper, pipeline-level view of structural bias .

Figure 2: Meta-evaluation of DiAlign on a balanced AmE–Ref(BrE) benchmark, showing the confusion matrix, classification performance, and confidence-margin distributions.

				Lexical Variants				DiAlign (broader variants)
Data Source	Document Type	Documents	Tokens	Orthographic		Vocabulary		% of
		(millions)	(billions)	AmE ( )	$\bm{\mathrm{Ref}}_{\bm{\mathrm{BrE}}}$ ( )	AmE ( )	$\bm{\mathrm{Ref}}_{\bm{\mathrm{BrE}}}$ ( )	AmE ( )	$\bm{\mathrm{Ref}}_{\bm{\mathrm{BrE}}}$ ( )
Book Corpus ( 2015 )	books	74	1.28	86.81	13.19	75.00	25.00	79.90 [0.79]	20.10 [0.76]
Wikipedia ( 2024 )	encyclopedic	6.4	4.3	72.94	27.06	61.43	38.57	72.18 [0.80]	27.82 [0.78]
Common Crawl (C4) ( 2020 )	web pages	365	156	75.12	24.88	67.00	33.00	73.67 [0.81]	26.33 [0.81]
Falcon RefinedWeb ( 2023 )	web pages	968	600	77.34	22.66	68.35	31.65	74.04 [0.82]	25.96 [0.75]

Table 1: Variant usage across six pretraining corpora. Under the Lexical Variants analysis, we report corpus-level distributions for orthographic (spelling-based) and vocabulary-based AmE–Ref(BrE) differences. We also report DiAlign scores, which capture broader regional alignment from grammatical, structural, collocational, and multi-word evidence; bracketed values denote mean alignment confidence. Datasets marked with * are evaluated on sampled subsets. RedPajama and Dolma contain mixed-domain content, including code, papers, forums, and social media. All probabilities are reported as percentages. Token statistics were computed using the LLaMA tokenizer ( Grattafiori et al., 2024 ) .

Figure 3: Distribution of AmE and Ref(BrE) variant probabilities across three pretraining corpora, shown separately for orthographic and vocabulary contrasts. Probabilities are computed from corpus frequencies over the matched variant pairs. Orthographic contrasts exhibit the strongest AmE skew, while vocabulary contrasts are more dispersed but lean toward AmE. Results for the remaining corpora are provided in Appendix Figure 7 .

Post-training stage	Hits	AmE (%)
Instruction tuning / SFT	16,348,769	75.95
Instruction tuning / chat	1,204,637	84.35
Instruction tuning / hum feedback	61,030	83.08
Instruction tuning / task mixture	6,578,816	71.77
RLHF / preference data	4,168,689	80.47
RLHF / reward-model feedback	77,332	84.72

Table 2: American English is consistently overrepresented across public post-training data. Results aggregate 21 datasets spanning instruction tuning, human feedback, preference, reward modeling, and safety. Variant hits counts matches to either form in the curated AmE–Ref(BrE) pairs, and AmE (%) reports the weighted AmE share. Dataset-level results in Appendix Table 9 .

				Orthographic			Vocabulary
Tokenizers	Developer Country	Model Access	Vocab Size	AmE ( )	$\bm{\mathrm{Ref}}_{\bm{\mathrm{BrE}}}$ ( )	$\Delta_{o}$	AmE ( )	$\bm{\mathrm{Ref}}_{\bm{\mathrm{BrE}}}$ ( )	$\Delta_{v}$
GPT-4	USA ( )		100K	2.73	2.86	$\uparrow$ 4.76 %	2.27	2.64	$\uparrow$ 16.30 %
GPT-4o	USA ( )		200K	2.65	2.77	$\uparrow$ 4.53 %	2.21	2.57	$\uparrow$ 16.29 %
Llama-3.3-70B	USA ( )		128K	2.72	2.85	$\uparrow$ 4.78 %	2.27	2.63	$\uparrow$ 15.86 %
Gemma-3-27B	USA ( )		262K	2.40	2.53	$\uparrow$ 5.42 %	2.02	2.35	$\uparrow$ 16.34 %
DeepSeek-V3	China ( )		128K	2.71	2.80	$\uparrow$ 3.32 %	2.37	2.67	$\uparrow$ 12.66 %

Table 3: Fertility scores for AmE and Ref(BrE) variants across the evaluated tokenizers, reported separately for orthographic and vocabulary contrasts. Lower fertility indicates more compact tokenization.

\Delta_{o}

and

\Delta_{v}

denote the relative fertility gap between matched AmE and Ref(BrE) forms. Bold and u nderlined values indicate the lowest and second-lowest fertility, respectively. All AmE–Ref(BrE) differences are statistically significant under the Wilcoxon signed-rank test (

p<0.01

), except for Velvet-2B in the orthographic category, marked with

\dagger

Figure 4: Tokenization granularity for $\mathrm{AmE}$ and $\mathrm{Ref}_{\mathrm{BrE}}$ variants across three tokenizers. Each panel shows the number of variants represented by 1, 2, or 3+ subwords. $\mathrm{Ref}_{\mathrm{BrE}}$ forms appear more frequently in the 3+ subword category, indicating greater long-tail fragmentation. Remaining tokenizers appear in Appendix Fig. 11 .

Figure 5: Semantic equivalence, unequal prediction cost. Top: matched

\mathrm{AmE}

–

\mathrm{Ref}_{\mathrm{BrE}}

counterfactuals are more similar than character-perturbed and unrelated controls. Bottom:

\mathrm{Ref}_{\mathrm{BrE}}

counterfactuals receive higher per-token loss across all five base/post-trained model families under equal sentence token counts. Error bars show paired-bootstrap 95% CIs.

		Natural Questions [ formal ]		ELI5 [ informal ]
LLMs	Developer Country	Default English ( $\text{AmE}^{\text{Default}}$ )	British English ( $\text{AmE}^{\text{BrE}}$ )	Default English ( $\text{AmE}^{\text{Default}}$ )	British English ( $\text{AmE}^{\text{BrE}}$ )
GPT-4o	USA ( )	79.00% [0.81]	45.33% [0.77]	77.00% [0.82]	3 4.67% [0.78]
Gemini-2.0-flash	USA ( )	76.00% [0.83]	51.00% [0.79]	75.33% [0.82]	42.33% [0.80]
Claude-3.7-sonnet	USA ( )	75.67% [0.85]	4 2.33% [0.80]	73.33% [0.86]	37.67% [0.78]
Llama-3.3-70B	USA ( )	74.67% [0.82]	47.33% [0.78]	69.00% [0.79]	30.00% [0.76]
Gemma-3-27B	USA ( )	69.33% [0.81]	45.67% [0.78]	6 8.33% [0.83]	38.00% [0.78]

Table 4: Regional preferences in LLM generation on Natural Questions ( formal ) and ELI5 ( informal ). We report the percentage of outputs classified as AmE under default English and British English (en-GB) prompting, with mean AmE alignment scores in brackets. AmE is the dominant default, while British-English prompting generally shifts outputs toward Ref(BrE). Bold marks the lowest AmE percentage in each column; u nderlined marks the second-lowest.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Difference Type	% of Pairs	Examples
	ends in “-or” (AmE) vs. “-our” (BrE)	2.26%	color ( ) vs. colour ( )	labor ( ) vs. labour ( )
	ends in “-ize” (AmE) vs. “-ise” (BrE)	11.58%	organize ( ) vs. organise ( )	realize ( ) vs. realise ( )
	ends in “-er” (AmE) vs. “-re” (BrE)	1.65%	center ( ) vs. centre ( )	liter ( ) vs. litre ( )
	ends in “-og” (AmE) vs. “-ogue” (BrE)	0.55%	dialog ( ) vs. dialogue ( )	catalog ( ) vs. catalogue ( )
	ends in “-ense” (AmE) vs. “-ence” (BrE)	0.22%	defense ( ) vs. defence ( )	pretense ( ) vs. pretence ( )
	“e” (AmE) vs. “ae” (BrE)	4.03%	esthetic ( ) vs. aesthetic ( )	pediatric ( ) vs. paediatric ( )

Appendix

Table 5: Distribution of the 1,813 matched AmE ( ) and Ref(BrE) ( ) variant pairs across linguistic categories in the curated resource. We report each category’s share of the full inventory with representative examples, grouped into orthographic ( spelling ) and vocabulary-based contrasts.

Level	American English (AmE) ( )	Reference: British English (BrE) ( )
Orthography	color, center, organize, traveling, favorite, check, jewelry, program, catalog, defense	colour, centre, organise, travelling, favourite, cheque, jewellery, programme, catalogue, defence
Vocabulary	cell phone, crosswalk, downtown, train station, parking lot, ZIP code, vacation, apartment, elevator, line, sidewalk, flashlight, soda, popsicle	mobile phone, zebra crossing, city centre, railway station, car park, postcode, holiday, flat, lift, queue, pavement, torch, fizzy drink, ice lolly
Grammar/usage	I just ate ; on the weekend ; Do you have a pen? ; The team is winning ; different than	I’ve just eaten ; at the weekend ; Have you got a pen? ; The team are winning ; different from
Conventions	double quotes; punctuation inside quotes; December 31, 2024 ; first floor ; 11:15 PM ; $68^{\circ}$ F	single quotes; punctuation outside quotes; 31 December 2024 ; ground floor ; 11.15 pm ; $20^{\circ}$ C

Appendix

Table 6: Representative distinctions between American English (AmE) ( ) and the British English reference, Ref(BrE) ( ), across orthography, vocabulary, grammar, and writing conventions. Examples reflect common majority-preference usage and are illustrative rather than exhaustive. Additional examples are provided in Tables 18 and 19 of the Appendix.

Figure 6: Illustrative walkthrough of DiAlign . Parallel AmE and Ref(BrE) passages contain matched contrasts in spelling, vocabulary, grammar, conventions, and multi-word usage. DiAlign decomposes each passage into overlapping 2–5-gram candidate spans, filters non-diagnostic spans, and compares the retained constructions against the American English (ID 17) and British English (ID 6) Google Books reference corpora. The resulting local evidence is aggregated into the regional alignment scores $(P_{\mathrm{AmE}},P_{\mathrm{Ref}_{\mathrm{BrE}}})$ .

Ablations	Accuracy	Precision	Recall	F1 Score	Avg. Conf. (AmE)	Avg. Conf. (BrE)
DiAlign (final)	93.18	90.67	96.25	93.38	0.84	0.91
– w/o Divergence Weight (DW)	92.25	89.29	95.98	92.52	0.77	0.85
– w/o Boosting Factor (BF)	92.25	89.29	95.98	92.52	0.82	0.89
– w/o Both (DW + BF)	91.31	88.13	95.45	91.65	0.75	0.83

Appendix

Table 7: Ablation study of DiAlign on the balanced AmE–Ref(BrE) meta-evaluation benchmark. We report classification performance (Accuracy, Precision, Recall, and F1) and mean confidence for AmE and Ref(BrE) predictions. Removing either divergence weighting (DW) or the boosting factor (BF) reduces performance, with the largest drop when both components are removed.

Reference period	Accuracy	Precision	Recall	F1
1950–2022	93.18	90.67	96.25	93.38
2000–2022	92.91	90.42	95.98	93.12

Appendix

Table 8: Temporal robustness of DiAlign to the Google Books reference window. Restricting the reference distributions from 1950–2022 to 2000–2022 changes accuracy by only 0.27 points and F1 by 0.26 points, indicating that performance is stable to the temporal range used for frequency estimation.

Figure 7: Variant-probability distributions for (a) BookCorpus, (b) Falcon RefinedWeb, and (c) RedPajama. Orthographic contrasts are more strongly concentrated toward AmE, while vocabulary contrasts are more dispersed but still lean toward AmE.

Figure 8: Category-level AmE and Ref(BrE) probabilities for (a) Wikipedia, (b) Common Crawl (C4), and (c) Dolma. Cells show the mean pair-level probability within each linguistic contrast category.

Figure 9: Category-level AmE and Ref(BrE) probabilities for (a) BookCorpus, (b) Falcon RefinedWeb, and (c) RedPajama. The same category-level pattern observed in Figure 8 is also evident across these corpora.

Figure 10: Source distributions in two pretraining corpora. (a) Top-level domains in C4.EN are concentrated in .com and .org , with much smaller shares from country-specific domains such as .co.uk . (b) High-frequency sources in Dolma illustrate the concentrated provenance of a curated mixed-domain corpus ( Raffel et al., 2020 ; Soldaini et al., 2024 ) .

Dataset	Stage	Rows	Hits	Weighted AmE (%)	Orth. AmE (%)	Vocab. AmE (%)	HF
Alpaca Cleaned ( Taori et al., 2023 )	SFT	51,760	74,991	85.35	94.84	80.05	HF
Databricks Dolly 15k ( Conover et al., 2023 )	SFT	15,011	19,997	65.29	78.09	67.65	HF
OpenAssistant Conversations v2 ( Köpf et al., 2023 )	SFT	61,278	61,030	83.08	92.56	76.85	HF
UltraChat 200k ( Ding et al., 2023 )	SFT	207,865	2,442,647	84.23	90.32	75.68	HF
OpenHermes 2.5 ( Teknium, 2023 )	SFT	1,001,551	2,517,105	81.81	85.78	74.01	HF
WizardLM Evol-Instruct v2 ( Xu et al., 2024 )	SFT	143,000	634,919	89.72	95.20	80.33	HF

Appendix

Table 9: Dataset-level post-training audit. Hits counts occurrences matching either member of the curated AmE–Ref(BrE) variant pairs. Weighted AmE (%) reports the aggregate AmE share across observed hits, while Orth. AmE (%) and Vocab. AmE (%) report mean pair-level AmE shares within orthographic and vocabulary categories, respectively. HF links identify the exact Hugging Face dataset releases used in the audit.

Figure 11: Additional tokenization-granularity results for (a) GPT-4, (b) Llama-3.3-70B, (c) StableLM-2-1.6B, (d) Mistral-Small-24B, (e) Velvet-2B, and (f) Falcon3-7B. Bars show the number of AmE and Ref(BrE) variants represented by 1, 2, or 3+ subwords.

Tokenizer	Developer Country	Tokenizer lineage	Independent	Interpretation
GPT-4	USA ( )	OpenAI cl100k_base	Yes	Independent US baseline
GPT-4o	USA ( )	OpenAI o200k_base	Yes	Independent US baseline
Llama-3.3-70B	USA ( )	Extended from US-origin tiktoken-style vocabulary	No	Derived; excluded from independent country-level evidence
Gemma-3-27B	USA ( )	Gemma/Gemini-family vocabulary	Yes	Independent US tokenizer
DeepSeek-V3	China ( )	DeepSeek BPE	Yes	Independent non-US tokenizer
Mistral-Small-24B	France ( )	Mistral Tekken family	Yes	Independent non-US tokenizer

Appendix

Table 10: Tokenizer provenance audit. Model developer country and tokenizer provenance are reported separately. Independent refers specifically to tokenizer-vocabulary provenance, not to model weights, architecture, or training data. Derived tokenizers are retained in the behavioral analysis but are not counted as independent country-level evidence.

Tokenizer compared with GPT-4	AmE	BrE	Validation	Generated	All
StableLM-2-1.6B	100.00	100.00	80.80	81.17	91.87
Llama-3.3-70B	99.17	99.06	99.87	99.83	99.43
GPT-4o	74.85	77.05	40.27	29.00	58.58
DeepSeek-V3	67.02	70.22	47.80	25.08	55.42
Mistral-Small-24B	65.47	66.91	26.47	10.00	46.11
Falcon3-7B	58.58	58.91	11.93	5.17	37.48

Appendix

Table 11: Tokenizer behavioral similarity with GPT-4. Values report the percentage of strings for which each tokenizer produces the same token-boundary sequence as GPT-4. StableLM-2 and Llama-3.3 show near-perfect agreement, in contrast to the substantially lower agreement observed for independently sourced tokenizer families.

Family	Stage	Rows	Gap / token	95% CI	$\mathrm{Ref}_{\mathrm{BrE}}$ higher
Llama	Base	1,008	0.189	[0.185, 0.193]	100.0%
Llama	Post	1,008	0.264	[0.262, 0.267]	100.0%
StableLM	Base	2,350	0.210	[0.204, 0.214]	100.0%
StableLM	Post	2,350	0.400	[0.389, 0.411]	100.0%
Ministral	Base	1,068	0.213	[0.208, 0.218]	100.0%
Ministral	Post	1,068	0.234	[0.229, 0.239]	100.0%

Appendix

Table 12: Prediction-cost asymmetry on the tokenizer-controlled RQ2 slice. Rows contain orthographic minimal-mention counterfactuals for which the

\mathrm{AmE}

and

\mathrm{Ref}_{\mathrm{BrE}}

sentences receive identical token counts under the evaluated model. Gap/token denotes the

\mathrm{Ref}_{\mathrm{BrE}}

minus

\mathrm{AmE}

difference in per-token negative log-likelihood, with positive values indicating higher prediction cost for

\mathrm{Ref}_{\mathrm{BrE}}

. Confidence intervals are paired-bootstrap 95% intervals.

Family	Stage	Hugging Face checkpoint
Gemma	Base	google/gemma-3-1b-pt
Gemma	Post-trained	google/gemma-3-1b-it
Ministral	Base	mistralai/Ministral-3-3B-Base-2512
Ministral	Post-trained	mistralai/Ministral-3-3B-Instruct-2512
Llama	Base	meta-llama/Llama-3.1-8B
Llama	Post-trained	meta-llama/Llama-3.1-8B-Instruct

Appendix

Table 13: Matched base and post-trained checkpoints used for the RQ2 prediction-cost analysis, spanning five model families and ten checkpoints.

Layer	Comparison	Models	Mean cosine
Embedding	$\mathrm{AmE}$ – $\mathrm{Ref}_{\mathrm{BrE}}$ counterfactual	10	0.902
Embedding	Character perturbation	10	0.824
Embedding	Unrelated control	10	0.504
Middle	$\mathrm{AmE}$ – $\mathrm{Ref}_{\mathrm{BrE}}$ counterfactual	10	1.000
Middle	Character perturbation	10	0.999
Middle	Unrelated control	10	0.997

Appendix

Table 14: Layerwise semantic-equivalence control. Values report model-level mean cosine similarities averaged across ten evaluated checkpoints. Matched

\mathrm{AmE}

–

\mathrm{Ref}_{\mathrm{BrE}}

counterfactuals are more similar than character-perturbed and unrelated controls across layers, with the clearest separation at the final layer.

Figure 12: Layerwise semantic-equivalence control. Matched

\mathrm{AmE}

–

\mathrm{Ref}_{\mathrm{BrE}}

counterfactuals remain closer than character-perturbed or unrelated controls across embedding, middle, and final representations, with the clearest separation at the final layer.

Panel of Figure 13 — Figure 13: Prompt used under the two language conditions. We set WORD_LIMIT $=50$ and {language} $\in$ { English , British English (en-GB) }.

Natural Questions (NQ) [formal]

ELI5 [informal]

Default English (

\%\text{AmE}^{\text{Default}}

)

British English (

\%\text{AmE}^{\text{BrE}}

)

Default English (

\%\text{AmE}^{\text{Default}}

)

British English (

\%\text{AmE}^{\text{BrE}}

)

LLMs

Model Access

#Words Avg. [SD]

Range [min-max]

#Words Avg. [SD]

Range [min-max]

#Words Avg. [SD]

Range [min-max]

#Words Avg. [SD]

Range [min-max]

GPT-4o

50.19 [1.75]

[46–57]

50.32 [1.61]

[45–56]

50.41 [1.60]

[47–55]

50.35 [1.53]

[46–55]

Gemini-2.0-flash

52.74 [2.71]

[46–60]

51.92 [2.89]

[45–61]

53.10 [2.62]

[47–61]

52.82 [2.95]

[44–60]

Claude-3.7-sonnet

45.25 [2.29]

[39–52]

45.49 [2.35]

[39–57]

47.70 [2.45]

[42–56]

47.38 [2.32]

[41–54]

Llama-3.3-70B

41.68 [7.12]

[16–50]

41.47 [7.44]

[18–50]

41.33 [7.83]

[21–50]

40.10 [8.48]

[18–51]

Appendix

Table 15: Word-length adherence across models in RQ3 (Section 5 ). Each model was instructed to produce exactly 50 words per response. We report mean length, standard deviation, and range for Natural Questions ( formal ) and ELI5 ( informal ) under default English and British English (en-GB) prompting.

Condition	Length bin	Avg. words	DiAlign AmE (%)
Default	1–50	44.08	67.64
Default	51–75	55.82	67.33
Default	76–100	86.70	64.31
Default	101–150	116.03	62.54
Default	151+	225.60	61.01
British-English steer	1–50	43.62	47.12

Appendix

Table 16: Realized response length and AmE alignment across length bins. Default-English outputs remain AmE-majority throughout, while British-English prompting produces the lowest AmE alignment among shorter, more instruction-adherent responses.

Figure 14: Timeline of independence across countries formerly under British colonization. The map highlights the wave of decolonization in the mid-twentieth century, when nations in Africa, Asia, the Caribbean, and the Pacific gained sovereignty. This geopolitical shift marked the decline of direct colonial governance but reinforced the institutional legacy of British English (BrE) in education, government, journalism, and law across many of these regions ( Tikly, 2016 ; Phillipson, 2018 ) .

Source	Title ( linked )	Description
Wikipedia	American and British English spelling differences	A widely cited reference outlining systematic orthographic differences between American and British English. The page provides examples of variant spellings (e.g., color vs. colour ), historical background, and explanations of regional conventions. It served as one of the authentic linguistic resources for curating consistent one-to-one variant pairs in our lexicon.
ThoughtCo.	American English to British English Vocabulary	A curated reference list of American and British English vocabulary differences, created by experienced educators and subject experts. Provides reliable lexical contrasts in an accessible format, supporting the construction of our AmE–BrE lexicon.
Research Article	Mapping the Americanization of English in Space and Time	An empirical study tracing how American English variants spread globally across regions and over time. Offers quantitative evidence of AmE–BrE lexical contrasts, providing authoritative grounding for the curated variant pairs in our unified lexicon.
IELTS	British vs. American English in the IELTS Test: Key Differences	An official IELTS guide highlighting key vocabulary, spelling, and grammar differences between AmE and BrE. The resource systematically documents contrasts across domains such as food, school, homes, and grammar, making it a practical reference for understanding standardized English variations.
Grammarly	How to Select Your English Dialect	A practical guide from Grammarly explaining how to switch between English dialects in writing tools, highlighting spelling, vocabulary, and usage variations (AmE vs BrE). Because it enumerates common dialectal choices in real writing, it serves as a useful supplementary resource for identifying variant pairs.
SpellZone	Sixty American English Words and their British English Counterparts	SpellZone provides a practical reference list of 60 common AmE–BrE word pairs, illustrating clear lexical contrasts in spelling and vocabulary. The resource highlights straightforward one-to-one mappings useful for systematic dialectal analysis.

Appendix

Table 17: Key linguistic and web-based sources used to construct the AmE–Ref(BrE) variant lexicon. Variant pairs were manually compiled across multiple sources, merged, and deduplicated to form a unified reference set for consistent analysis.

Category	British English (BrE)	American English (AmE)
o vs. ou	colour, honour, behaviour	color, honor, behavior
-re vs. -er endings	centre, fibre, theatre	center, fiber, theater
-ise vs. -ize endings	recognise, authorise	recognize, authorize
-yse vs. -yze endings	analyse, paralyse, catalyse	analyze, paralyze, catalyze
Single vs. double l (inflection)	travelled, counselled	traveled, counseled
-ll + -ly suffix	skilfully, wilfully	skillfully, willfully

Appendix

Table 18: Representative AmE–Ref(BrE) distinctions across orthography, grammar, and formatting conventions. Entries reflect common majority-preference usage and are illustrative rather than exhaustive.

Category	British English (BrE)	American English (AmE)
Preposition before days	She resigned on Thursday.	She resigned Thursday.
Street naming	in the High Street	on Main Street
Transitivity (protest)	protest against discrimination	protest discrimination
Ditransitives (write)	write to me	write me
Meeting collocation	meet the team	meet with the team
Transport & wayfinding

Appendix

Table 19: Representative AmE–Ref(BrE) preferences across everyday domains, including transport, household, and food. Entries reflect common majority-preference usage and are illustrative rather than exhaustive.

Explore similar work

May 15, 2026cs.CL

Toward LLMs Beyond English-Centric Development

Through an analysis of sequences generated by open-weight large language models (LLMs), we demonstrate that LLMs are heavily biased toward English. While continual pre-training is commonly used to adapt LLMs to a target language, we show that it does not offer a cost advantage over training from scratch, even for improving cultural understanding in the target language. These findings suggest that dedicated per-language investment may become increasingly important for future LLM development, rather than relying primarily on the expansion of English-centric resources.

Sho Takase, Ukyo Honda

CyberAgent

Large Language Model Bias Pretraining

Ask PDF

Jun 1, 2026cs.CL

Multilinguality of Large Language Models From a Structural Perspective

Large language models (LLMs) have excelled in processing multiple languages through pre- and post-training on multilingual data, even though English dominates the training data. Prior work focusing on token representations has revealed how those LLMs process non-English text. Although these analyses have provided insightful findings, they fail to capture a structural view, which is an inherent property of language. In this study, we explore the multilinguality of LLMs through representational structural analysis. Our findings reveal that low-resource languages are structurally more different from English than high- and mid-resource languages, and that language-specific post-training alters their structures while preserving inter-language relationships.

Haruki Sakajo, Yusuke Sakai, Hidetaka Kamigaito +1

Nara Institute of Science and Technology (NAIST)

Multilingual Language Models Multilingual Low-Resource Languages Representational Capacity

Ask PDF

Jul 7, 2026cs.CL

LLMs Silently Correct African American English: Auditing and Mitigating Dialect Bias via Activation Steering

African American English (AAE), a rule-governed dialect spoken by over 30 million people, is routinely misinterpreted and "corrected" by large language models (LLMs). Across six instruction-tuned LLMs (14B to 70B), we show that state-of-the-art models systematically prefer Standard American English (SAE) continuations even when the preceding context is in AAE, effectively rewriting AAE into SAE. We present an end-to-end framework to audit and mitigate this bias. For auditing, we introduce conditional Dialect Group Invariance (cDGI), which isolates true model bias from translator-induced artifacts, and a feature-level localization analysis that identifies which AAE markers most strongly trigger bias; we find that syntactic constructions, especially negative concord (e.g., "ain't nobody"), are universal triggers across all models. For mitigation, we introduce, to our knowledge, the first application of activation steering to dialect bias: a training-free, test-time method that extracts dialect directions via causal tracing and injects them into bias-relevant layers. Activation steering reduces bias 5 to 20 times more than prompting while preserving SAE fluency. To enable this work, we release REAL-AAE , the largest real-AAE parallel corpus to date: 17,479 AAE/SAE/ AAE_back triplets from natural tweets (2 to 6 times larger than prior real-AAE resources), validated automatically (BERTScore F1 = 0.95) and by three native AAE speakers (83.0% semantic agreement).

Huan Wu, Ali Emami, Muhammad Furquan Hassan +5

1York University · 2Vector Institute · Connected Minds +6

Large Language Model Bias Dialects Linguistics Model Auditing

Ask PDF