Leave a Comment
Your email address will not be published. Required fields are marked *
AI Writes in English: But Does It Understand Non-Native English Speakers?
This scene plays out millions of times a day, and it raises a question the AI industry has largely avoided asking: when AI writes in English, does it actually understand the people who write English as a second, third, or fourth language-or does it merely correct them?
The hidden assumption: "better" means "more native-like"
Many AI writing tools appear to operate on a quiet assumption: that better English means English that sounds as if an educated native speaker of a dominant variety had written it. This is my reading of how the tools behave, not a documented design policy.This assumption is baked into training data, into what models treat as "high quality," and into interfaces that offer one-click "improvements" without explaining what is being improved, or toward what.
Applied linguistics has a related concept. Adrian Holliday described 'native-speakerism' as an ideology in English language teaching that treats 'native-speaker' teachers as the source of the ideals of English and of its teaching. Applying it to AI writing tools extends the concept, but it is a natural extension. Scholars of Global Englishes have long challenged this hierarchy, pointing out that English is now a pluricentric language with many legitimate centers-Indian, Nigerian, Singaporean, and other varieties included-and that English as a lingua franca belongs to its speakers, not to its so-called owners (Cambridge Language Teaching research agenda). As Barbara Seidlhofer argued in the ELT Journal, since non-native speakers now vastly outnumber native speakers, it can no longer be claimed that English is the property of its native speakers (ELT Journal). Recent work extends this critique directly to education, arguing that the persistence of native-speaker norms impoverishes rather than protects communication (Global Englishes review).
Ethnologue estimates that English has about 1.5 billion users, and most of them use it as an additional language. Yet the typical AI assistant treats these users not as multilingual communicators with their own valid Englishes, but as people whose deviations from Anglo-American norms are defects awaiting repair.
The problem with sounding native
Here it helps to separate concepts that AI tools routinely blur:
- Correctness – conforming to the grammatical rules of a described variety.
- Clarity – being understood.
- Fluency – producing language smoothly.
- Style – rhetorical and tonal choices.
- Identity – the linguistic fingerprints of culture, personality, and history.
Native-like English guarantees none of the others. A sentence can be perfectly native and unclear, or grammatically unremarkable by Anglo-American standards and completely effective. When an AI tool conflates these dimensions, it does not just mislabel a sentence-it misleads a writer about what the problem actually is.
Research on generative AI's effect on text should give us pause here. Analyses of academic abstracts show that ChatGPT-assisted writing tends to converge on a distinctive "LLM vocabulary"-certain words spike while the overall stylistic range narrows (arXiv: Is ChatGPT Transforming Academics' Writing Style?). A study presented at ICML 2024 estimated that 6.5% to 16.9% of peer-review text at several AI conferences could have been substantially modified by large language models, and found that higher estimated modification correlated with more homogeneous review content(Monitoring AI-Modified Content at Scale). In a Science Advances experiment, writers who received story ideas from a language model produced short stories that evaluators rated as more creative, yet the AI-assisted stories were more similar to one another(Science Advances).
None of these studies examines multilingual researchers directly, so applying them to multilingual writers is my inference. A 2026 preprint proposes the term 'cultural ghosting' for what it measured: it found that, in Indian, Nigerian and Singaporean English texts processed by five models, culturally marked expressions were erased about 10% of the time overall, and politeness markers far more often than vocabulary(arXiv: Cultural Ghosting). The message survives; the messenger fades.
The multilingual user's dilemma
Consider three everyday cases:
- Academic writing. A graduate student writes, "This study tries to show how community knowledge matters." The tool offers: "This study demonstrates the significance of community-based epistemologies." The revision is more formal-and more abstract, more distant from what the student meant, and indistinguishable from a thousand other abstracts.
- Professional email. An entrepreneur writes, "I would be very grateful if you could share your thoughts when convenient." The AI rewrites: "Please share your feedback." Shorter? Yes. But the politeness strategy-a core part of how many cultures build trust-has been deleted, and no error was ever corrected.
- Everyday messaging. "Let's meet tomorrow if it's okay with you" becomes "Meeting tomorrow is confirmed." Efficient, colder, and no longer really the writer's.
None of these rewrites fixed a mistake. Each replaced a choice with a default. And because tools rarely distinguish "ungrammatical" from "unidiomatic in Standard American English" from "less formal," the writer learns nothing except that their instinct was wrong.
The stakes can be higher than style. One widely discussed 2023 study in Patterns tested seven GPT detectors on 91 human-written TOEFL essays and 88 essays by US eighth-graders. The detectors misclassified the TOEFL essays as AI-generated about 61% of the time on average, and far more rarely misclassified the US students' essays. Because the essays were human-written, these were false positives, not evidence of AI used.The study's authors warned against using detectors in evaluative settings, particularly for non-native speakers, and speculated that non-native writers might turn to AI tools to sound more native.Recent HEPI blog posts argue that AI detection raises fairness concerns for international students in UK universities.
Multilingual writers are thus caught in a double bind: use AI to sound native and risk accusations (and voice loss), or refuse it and be judged against a native standard that institutions, and sometimes the tools, may reinforce.
From correction to empowerment
This is why I want to propose a different framing-call it AI language empowerment. The goal of a writing tool should not be to make a multilingual person sound like a native speaker. It should be to help them become a more effective communicator while keeping control of their voice, meaning, identity, and learning.
The distinction matters because correction and empowerment are different philosophies, not different feature lists:
- Correction asks: how do I make this text match the standard?
- Empowerment asks: what does this writer want to achieve, and what do they need from me to achieve it?
A corrected text is the tool's output. An empowered writer is the tool's outcome. Research on AI and language learning shows the second path is realistic: controlled studies find LLM-based chatbots significantly aid second-language vocabulary development (peer-reviewed study), and emerging work on AI feedback for ESL academic writing shows large gains in grammar, vocabulary, and cohesion (Springer: Beyond Red Ink). Early evidence suggests AI can support second-language writers. My argument is that supporting them as writers, not just correcting their text, does not appear to be the default design goal of most tools.
What a different design would look like
Imagine a tool built on empowerment. Not features bolted onto a rewriter, but expressions of a single ethic: the writer is the author; the AI is the instrument.
"Your sentence is grammatically incorrect" versus "Your sentence is grammatically acceptable, but this alternative is more common in academic English" versus "Your sentence is understandable; this change would make it more formal." These are three different statements with different truth values. A tool that cannot tell them apart is not helping you communicate; it is training you to conform. Designers of human-AI writing systems are actively studying how to preserve a sense of authorship precisely because wholesale rewriting weakens it (Who Owns the Text?, arXiv).
Preserve my voice. Let users lock in tone, register, and stylistic markers-indirect requests as politeness, longer relational greetings, a particular rhythm-so the AI works within their voice rather than over it. This flips the default: transformation on request, not by default.
Explain rather than replace. Every suggestion comes with a reason and a category: error, idiom, register, variety. Research on Grammarly-style feedback suggests users often accept suggestions readily. My concern is that acceptance without understanding may build dependence rather than competence(study of native vs. non-native Grammarly use).
Multiple Englishes. The tool recognizes Indian, Nigerian, Singaporean, and other Englishes as standards in their own right, not as error lists. There is precedent for culture-aware behavior: A COLING 2025 study found weak but significant shifts in GPT-4o's questionnaire responses across languages and across five English-speaking countries, which the authors link to stereotypes-which cuts both ways, showing models already encode cultural signal, just not transparently (Open University of Catalonia).
Language-learning partner mode. The tool tracks recurring patterns and offers targeted micro-lessons-"you consistently omit articles here's the rule"-turning each correction into durable learning. Studies of multilingual co-writing tools built for specific linguistic communities, like a Tamil-English co-writer for first-language support, show this inclusive design is technically feasible (Tamil Co-Writer, LAK 2024 workshops). Research on multilingual co-writing also shows users' choices shift depending on whether the AI engages their full linguistic repertoire, not just the English endpoint (Mind the Gap!, arXiv).
Confidence-preserving corrections. Name what works before what deviates. For many multilingual professionals, writing in English is an act of sustained courage; a tool that only ever highlights failure quietly teaches them to stop trusting their own judgment.
Why this matters beyond English
Extend the logic outward. In research, scholars debate whether non-native speakers face structural "linguistic injustice" in publishing; Ken Hyland has argued that the 'linguistic injustice' framing is overstated and has called for direct evidence; Politzer-Ahles and colleagues responded that the disadvantage is real. The debate is open, but the pressures it describes are documented in other work(Hyland's "myth of linguistic injustice" paper; Politzer-Ahles's response). A commentary published by the American Society for Microbiology discusses the hidden costs of linguistic bias for non-native English-speaking researchers(American Society for Microbiology). In education, detection bias turns language support into accusation (HEPI).
AI tools now sit at this chokepoint. At very large scale, they can quietly reinforce a single standard of English, or they can be designed with the multilingual majority in mind.
The counterargument, and a response
Let me be fair to the other side. Many multilingual users genuinely want native-like English, and who is anyone to refuse them? In high-stakes contexts-law, medicine, international publication-convergence on a standard genuinely reduces friction, and AI assistance can improve L2 writers' accuracy and fluency (Springer: Beyond Red Ink). Standardization is sometimes exactly the service being purchased.
And yes, some users want to sound native precisely because gatekeepers-editors, employers, reviewers-reward it. Blaming the tool alone is too easy.
But this concedes my point rather than refuting it. The issue is not whether native-like English is a legitimate goal; it is whether it should be the unspoken default for every user, every time, with no alternatives offered and no distinctions explained. A tool that assumes the goal imposes it. A tool that asks-Native polish? Keep my voice? Teach me the rule?respects the writer. Offering that control may cost little; withholding it risks costing writers their voice by default.
Conclusion
AI writing tools may be the most widely used writing technology multilingual researchers have had. The question is what it does with that power: continue treating the world's English speakers as imperfect approximations of a Michigan or London ideal, or recognize them as what they are-the majority of English's users, and increasingly its future.
The next time a tool offers to "improve" your English, ask it one question first: improve it for whom, toward what, and at what cost to the person writing?
Because the real question was never whether AI can write perfect English. It is whether AI can help you write your English-clearly, confidently, and unmistakably as yourself.
Keywords
AI Writing Tools Multilingual Researchers Language Equity Non-Native English Speakers Scholarly Communication Linguistic Diversity AI Language Empowerment AI Detection Bias Academic Writing Research Communication
Aisha Abdullahi Mahmud is a doctoral researcher in plant biotechnology at the Technical University of Munich, working on plant–microbe interactions and bioinformatics-driven approaches to crop resilience. A multilingual scientist writing at the intersection of language and technology, she believes AI should amplify voices, not standardize them.
View All Posts by Aisha Abdullahi MahmudDisclaimer
The views and opinions expressed in this article are those of the author(s) and do not necessarily reflect the official policy or position of their affiliated institutions, the Asian Council of Science Editors (ACSE), or the Editor’s Café editorial team.
