Publishing in English From a Non-Dominant Language

For most researchers on the planet, the hardest part of publishing a paper is not the research. It is writing it up in a language they did not grow up speaking. English runs the journals, and a writer working from Arabic, Mandarin, Spanish, or Polish pays a tax that a native speaker never sees: the same finding, less fluently phrased, reads as weaker work. Generative AI erases that tax almost overnight — and in doing so it moves the barrier rather than removing it. The old barrier was can you write it in English. The new one is can you prove the ideas are yours, and did you say how you got the words. That is the line careers now break on, and it is far less obvious than the language gap it replaced.

This piece is about where that line falls: what you are allowed to use, what you have to disclose, and the narrow zone in between where honest writers get into trouble without meaning to. It sits inside the larger question of what it costs to write as a non-native English speaker in a system built around native norms.

The barrier is real, and it was measured before AI arrived

Start with the thing AI is supposedly fixing, because it helps to know its actual size. The expectation that academic English should sound native is not imagined. Abdulaziz Alfehaid and Nada Alkhatib document the ideology behind it, quoting Trudgill’s blunt claim that “the true repository of the English language is its native speakers,” and noting that non-conformity to native norms has historically been “often criticized and even penalized” (Alfehaid & Alkhatib, 2023, p. 4). A writer whose English carries the fingerprints of their first language — a preposition off, a rhythm that isn’t quite idiomatic — has long risked being read as less rigorous, not just less fluent.

Ziyang Xu’s analysis of how non-native scholars actually behave shows the workaround already in full swing before ChatGPT dominated. Examining 168 AI-usage declarations pulled from 8,859 papers, Xu found non-native authors leaning far more heavily than native authors on tools like DeepL and Grammarly for translation and grammar checking, with the difference between groups statistically significant (Fisher-Freeman-Halton test, p = 0.0012) (Xu, 2025). The higher Grammarly use among non-native groups, Xu argues, reflects their sharper need for language accuracy — they are buying down the tax. And the newer tools promise to buy it down completely: Hengzhi Hu and colleagues note that ChatGPT is pitched to non-native researchers as a way “to mitigate linguistic disadvantage by improving grammatical accuracy, coherence, and fluency,” thereby “leveling the playing field in the global research arena” (Hu et al., 2025, p. 2).

That promise is real. The catch is that leveling the field with a machine raises a question the field never had to answer before: whose writing is this?

The norm shifted — but not in the way the tools claim

Here is a fact that complicates the whole “AI levels the playing field” story: the native-like standard was already softening on its own. When Alfehaid and Alkhatib interviewed 33 lecturers at UK universities about whether postgraduates must write in native-like English, only four said yes (Alfehaid & Alkhatib, 2023, p. 9). Their overall finding was that “native-like English seems to not be required at all by many departments,” and that “total conformity to native English norms was not required at all” (Alfehaid & Alkhatib, 2023, p. 12). What departments actually demand is that writing be “clear, coherent, and comprehensible” (Alfehaid & Alkhatib, 2023, p. 12) — a much lower and more humane bar than sounding like you were born in Manchester.

This matters for the disclosure question, because it changes what you are reaching for AI to do. If the real standard is comprehensible, then a light grammar pass to fix agreement and articles is enough — and, as we’ll see, that kind of help is the least fraught to use and the least likely to need declaring. If you believe the standard is native-fluent, you push the tool further, into rewriting and restructuring, and you cross into the zone that publishers now treat as content creation. A lot of non-native writers over-reach on AI because they are still chasing a native standard that their own examiners have quietly abandoned. What AI feedback reliably improves — and where it stops — is worth knowing in detail, because it maps almost exactly onto the disclosure line; we cover it in what AI feedback actually improves in L2 writing.

What you’re allowed to use

The good news is that the permitted zone is wide, and it covers most of what a non-native writer actually needs. Across the major publishers a rough consensus now holds: basic grammar, spelling, and punctuation checking is treated as ordinary tooling and is generally exempt from disclosure — the same category as a spellchecker (Enago, 2025). Nobody expects you to declare that you used autocorrect, and a light Grammarly-style pass sits in the same bucket at most journals.

The line moves as the help gets heavier. Elsevier, for instance, permits using AI to improve the language of a manuscript under human oversight, but requires that use to be stated in a declaration — and the same policy exempts basic grammar and spelling (Enago, 2025). The pattern across publishers is a gradient: mechanical correction sits below the disclosure threshold; language improvement, rephrasing, and translation increasingly sit above it, because publishers now read those as shaping content rather than just cleaning it (Enago, 2025). Xu’s data captures the moment this became formal: publishers like Elsevier introduced AI-usage disclosure templates that require authors to specify which tools they used and for what purpose (Xu, 2025).

So the allowed set, in practice:

• Grammar, spelling, punctuation correction — use freely, disclosure rarely required.
• Language editing and rephrasing for fluency — usually permitted, usually must be disclosed.
• Translation from your dominant language into English — permitted at most journals, but treated as substantive and must be disclosed.
• Generating text, arguments, or references from a prompt — this is where “assistance” ends and authorship problems begin.

What you must disclose — and who says so

The one thing every major body agrees on is that AI cannot be an author. The Committee on Publication Ethics, the World Association of Medical Editors, and the ICMJE have all held since 2023 that authorship requires accountability for the work, and a tool cannot be accountable (COPE, 2023). What flows from that is the disclosure rule: where AI is used in drafting, editing, translation, or summarizing, you state which tool you used and how, typically in the methods or acknowledgments, and you remain fully responsible for everything it produced (COPE, 2023; Enago, 2025).

The gap between policy and behavior is the current live problem. A 2025 analysis found that journal AI policies are failing to curb the surge in AI-assisted writing — the rules exist, but undisclosed use is widespread — which is precisely why editors are getting stricter about enforcement rather than looser. For a non-native writer, that enforcement climate is dangerous in a specific way, because the tools built to catch undisclosed AI are the same tools that systematically misread second-language prose.

The trap: where honest non-native writers get burned

Two hazards sit right on the line, and both fall harder on non-native authors.

The first is detection. AI detectors flag non-native English at rates that would be a scandal in any other measurement context — Doğan and Doğan cite controlled evaluations with false-positive rates exceeding 60% on genuine non-native academic writing (Doğan & Doğan, 2026). Worse, these accusations are almost impossible to disprove: unlike plagiarism, AI authorship “lacks verifiable textual overlap, rendering accusations inherently ambiguous” (Doğan & Doğan, 2026). The authors’ recommendation is that where detectors are used at all, their role must be disclosed and authors must be allowed to contest the finding (Doğan & Doğan, 2026). If you write clean, careful English as a second language, you can be accused of using AI whether or not you did — a mechanism we unpack in why detectors flag non-native English writers.

The second hazard is the one that ends papers: fabricated references. Hu and colleagues note that a growing number of publications from non-native English researchers “display signs of AI involvement — ranging from stylistic uniformity to the inclusion of fabricated or unverifiable references” (Hu et al., 2025, p. 2). This is the failure mode that turns permitted assistance into a retraction. A translation tool will not invent a citation; a generative model asked to “add supporting sources” absolutely will, and it will format the fake convincingly. If you use AI anywhere near your bibliography, verify that every source it touched actually exists before you submit — you can run the reference list through a free bibliography checker to catch invented or mismatched citations before an editor does. A disclosed AI language edit is a routine thing. A single hallucinated reference is a misconduct finding.

The identity cost the policies don’t mention

There is a quieter tax that no disclosure form captures. Hu and colleagues interviewed 25 non-native researchers, most of whom had already published in peer-reviewed journals and edited books using ChatGPT (Hu et al., 2025, p. 4), and found five distinct ways they made peace with the tool. The first they call “reluctant adoption” — early use “marked by secrecy and moral tension” (Hu et al., 2025, p. 1). These are competent scholars who felt they were doing something faintly shameful by leveling their own field.

That secrecy is exactly what good disclosure norms are meant to dissolve. If using AI to bring your English up to clear and comprehensible is permitted and declared, there is nothing to hide, and the moral tension has nowhere to live. The writers who get hurt are the ones who treat permitted, disclosable help as a secret — because a secret, once discovered, looks like the thing it was never allowed to be.

Where the line falls, in one paragraph

If you write in English from a non-dominant language, use AI to reach comprehensible, not native — that lower bar is the one your readers and examiners actually hold you to (Alfehaid & Alkhatib, 2023, p. 12). Grammar and spelling correction is yours to use freely. Language editing, rephrasing, and especially translation are permitted at most journals but must be disclosed — name the tool and the purpose, and keep it out of the shadows (COPE, 2023; Xu, 2025). Never let a generative model near your citations without verifying every one. And hold onto the ideas: the disclosure rules exist to protect the one thing that was always yours — the argument — while letting a machine carry the words across a border you were never obliged to cross alone.