Prompting experiments show LLMs perform badly at generating non-US national standard varieties and conflate Nigerian English with Nigerian Pidgin¶

Claim · All claims
Evidence · 1 study (2 entries) · qualitative r2 · q1
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q1 i? In the contributors' prompting experiments, LLMs performed very badly at generating national standard varieties of English other than US English, even failing to implement British spelling consistently. → Ugwuanyi 2025 q1 i? ChatGPT's one-shot rewrite in 'Nigerian English' produced text with exaggerated Nigerian Pidgin features that would no longer be considered academic, indicating minoritised varieties are underrepresented in pre-training data. → Ugwuanyi 2025 (2)

Evidence¶

Ugwuanyi 2025¶

Ugwuanyi, K., Mair, C., Dovchin, S., Erdocia, I., Kuteeva, M., & Airemionkhale, E. (2025). Generative AI and linguistic diversity in academic writing and publishing: Perspectives from World Englishes. https://arxiv.org/abs/2607.28505 doi:10.48550/arxiv.2607.28505

q1 · i? · theoretical · r2

Self-reported prompting experiments by a World Englishes scholar, reported in the dialogue: LLMs "perform very badly" at generating non-US national standard varieties, including inconsistent British spelling over more than a few paragraphs. Informal, unreported methodology.

"I have carried out prompting experiments to generate text in national standard varieties of English other than the US one and found LLMs to perform very badly at this task."

Ugwuanyi 2025 (2)¶

Ugwuanyi, K., Mair, C., Dovchin, S., Erdocia, I., Kuteeva, M., & Airemionkhale, E. (2025). Generative AI and linguistic diversity in academic writing and publishing: Perspectives from World Englishes. https://arxiv.org/abs/2607.28505. doi:10.48550/arxiv.2607.28505

q1 · i? · qualitative · r2

Analysis of a ChatGPT-4 one-shot prompt rewriting an abstract in Nigerian English: the output exaggerated enregistered Pidgin-style features, showing minoritised varieties are "underrepresented in the pre-training data". Single-prompt demonstration, not a systematic evaluation.

"the above text generated by ChatGPT confirms that the voices of scholars using minoriti sed varieties of World Englishes are clearly underrepresented in the pre-training data."

Discussion¶

Related Claims¶

• In LLM-assisted team formation, persona nationality systematically shifts assignments, with Nigerian names favored for Quality Assurance and South Korean names for Interface Design relative to Chinese names — related
• Current LLMs reproduce linguistic hierarchies by being built on dominant standard English varieties, producing a monolithic English that ignores regional variation — a broader claim this one bears on
• Native speakers rated outputs of LLMs fine-tuned on Nigerian Pidgin English as understandable and context-aware, though no non-fine-tuned baseline was evaluated — related
• Automated grading systems may produce systematically inequitable outcomes for international and non-Western students — related