Artificial intelligence tools (AI tools) can now draft paragraphs, polish grammar, summarise papers, and suggest structure in seconds, and most researchers have tried at least one of them. Used well, these tools can help writers — especially those writing in a second language — express their own ideas more clearly and spend more time on thinking than on phrasing. Used carelessly, they can produce confident-sounding text that is wrong, copied, or fabricated, and they can put authors in breach of journal policies they never read.
This guide gives a balanced, practical account of how to use AI tools in research writing without crossing ethical lines. It covers what AI tools can and cannot do, what disclosure and authorship rules generally expect, the accuracy and plagiarism risks, and a workflow that keeps the author firmly in control. Journal policies on AI differ from journal to journal and change over time, so always check your target journal's current author guidelines before submitting — this guide explains the general principles, not any one journal's specific policy. It connects with our guides on checking and avoiding plagiarism and research ethics, because AI use raises questions in both areas.
AI writing tools are strongest at language tasks: improving clarity, fixing grammar, varying sentence structure, and helping a writer organise existing ideas. They are also useful for brainstorming, outlining, and summarising material you have already read. For these uses, the author remains the source of the ideas, and the tool functions much like an advanced language editor.
They are weakest exactly where research writing matters most: facts, citations, and reasoning. AI tools can invent references that look real but do not exist, quote statistics that were never published, and produce plausible-sounding explanations that are wrong. They do not know what is true; they predict what text is likely to follow. That limitation shapes every rule that follows: anything an AI tool produces must be treated as an unverified draft, not as a source, and every factual claim, reference, and number must be checked against a real source before it enters a manuscript.
Most major publishers and guidance bodies now expect authors to disclose AI use in their manuscripts. The usual expectation is a statement, often in the methods or acknowledgements section or in a dedicated declaration, describing which tool was used and for what purpose — for example, that an AI tool was used to improve language clarity, while all ideas, data, and conclusions are the authors' own. Disclosure is not an admission of wrongdoing; it is transparency, and reviewers and editors increasingly treat undisclosed AI use as a problem in itself.
Because expectations vary, check three places before submitting: the journal's author guidelines, the publisher's AI policy, and any funder or institutional policy that applies to your work. If the journal requires a specific disclosure wording or location, follow it exactly. If you used no AI tool at all, there is nothing to disclose — but if you used one even for editing, disclose it where the journal asks. When in doubt, disclose more rather than less: an editor can ignore an unnecessary disclosure, but cannot forgive one that was hidden.
The consistent position across major publication guidance is that AI tools cannot be listed as authors, because authorship requires accountability: an author must be able to take responsibility for the work, answer for its accuracy, and agree to its publication, and a tool cannot do any of those things. Responsibility stays with the human authors for every word in the manuscript, including text an AI tool drafted. That principle is worth stating plainly, because it answers most ethical questions about AI use: if you would not sign your name to a claim, do not let an AI tool put it in your paper.
The same principle applies to peer review and to unpublished material. Do not paste manuscripts under review, confidential data, or unpublished findings into public AI tools, because the material may be stored or used beyond your control. For the same reason, be careful with sensitive data — participant information, clinical details, or commercially sensitive material should not be entered into tools whose privacy terms you have not checked. When your institution provides an approved tool with appropriate protections, prefer it over a public one for sensitive work.
Three risks deserve separate attention, because each has caught out real authors. The first is fabrication: AI tools invent citations, quotations, and data with complete confidence. Every reference an AI tool suggests must be verified in the original source, and any reference you cannot verify must be removed, however well it fits your argument. The second is plagiarism: AI-generated text can reproduce phrases from its training material, and it can also paraphrase so closely that the result is still effectively copied. Run your manuscript through plagiarism checks as usual, and rewrite any passage that depends on wording you did not originate — our guide on plagiarism checking explains how.
The third risk is subtler: the erosion of your own thinking. A manuscript assembled from AI suggestions can read smoothly while saying very little, because the tool supplies structure without understanding. Readers and reviewers notice this quickly — papers that are fluent but empty, with generic claims and no specific evidence, are a growing cause of rejection. The test is simple: remove the AI tool and ask whether you can still defend every claim in the paper from your own knowledge and sources. If not, the tool has done too much of the thinking, and the manuscript needs more of your own work.
A workflow that keeps you in control has a clear order. First, do the research and the thinking yourself: the question, the gap, the design, and the interpretation must be yours, informed by sources you have actually read. Our guides on finding a research gap, writing a literature review, and writing a methodology section cover those foundations. Second, draft in your own words, even roughly — a rough draft you understand is worth more than a polished draft you cannot defend.
Third, use AI tools for defined, limited tasks: clarifying a paragraph you wrote, checking grammar, generating an outline from your own notes, or summarising a paper you have read and will verify. Fourth, verify everything the tool produced: every reference, every number, every factual claim, against a real source. Fifth, disclose the use as your journal requires, and keep a note of what you used for what, in case an editor asks. Finally, read the whole manuscript aloud before submission. If a sentence sounds like something you would not say, or a claim sounds like something you cannot support, rewrite it yourself. That final read is the cheapest quality check in academic writing, and it catches more AI problems than any other step.
The most serious mistake is submitting AI-generated text without checking it, especially references and data — fabricated citations are now one of the easiest problems for editors to detect, and they destroy trust in the whole paper. The second is failing to disclose AI use where the journal requires it. The third is entering confidential or sensitive material into public tools. The fourth is letting AI tools write the interpretation: your discussion should reflect your judgement about what your results mean, not a tool's generic summary. And the fifth is over-reliance: if every section of a paper was drafted by a tool, the paper usually reads as generic, and reviewers will say so. AI tools are best treated as assistants for language and organisation, not as substitutes for research.
Used that way, they are genuinely helpful — particularly for researchers writing in English as an additional language, for whom polishing support can make the difference between a good study being understood and being dismissed for its prose. The ethical line is not whether you used a tool, but whether you remained the author: your ideas, your evidence, your verification, your responsibility. Our guide on how to publish a research paper covers the submission process where disclosure statements are made, and our guide on writing a research abstract helps with the summary that AI tools should never be left to write alone.
For researchers writing in English as an additional language, AI tools offer a genuine benefit that deserves honest acknowledgement: they can turn a well-designed study, clearly thought through, into prose that reviewers judge on its science rather than its grammar. Many strong studies from non-English-speaking countries have struggled in review not because the research was weak, but because the writing obscured it, and a tool that helps an author express their own ideas more clearly serves the whole research community. Used for this purpose — polishing sentences the author wrote, suggesting clearer phrasing for ideas the author already has — AI assistance is close to what human language editors have always done, and it should be disclosed in the same spirit.
The risks, however, are also higher for second-language writers, and they deserve plain statement. AI tools trained mostly on English text can flatten an author's voice into generic phrasing, replacing precise disciplinary terms with vaguer ones, or "correcting" a sentence into something that no longer means what the author intended. They can also introduce errors that a native speaker might catch instinctively but a second-language writer might miss, such as a changed preposition that alters a result's meaning, or a synonym that is wrong in a technical context. The safeguard is the same as for every other use: read the edited text against your own meaning, sentence by sentence, and keep any phrasing you are unsure about in your own words rather than the tool's.
A practical approach is to draft first in whatever way lets you think most clearly — including drafting sections in your first language and translating them yourself — and then use the AI tool only on the English draft, for clarity and grammar. Keep a record of what the tool changed, compare the edited version with your draft, and ask a colleague fluent in English to read the final manuscript if one is available. Disclosure wording can be simple and factual, for example: "An AI tool was used to improve the clarity of the English language in this manuscript; all ideas, data, and conclusions are the authors' own." That sentence protects you, informs the editor, and accurately describes assistance that helped the science be understood.
| Use | Risk level | What to do |
|---|---|---|
| Grammar and clarity editing of your own draft | Lower | Disclose where required; check the edited meaning is unchanged |
| Outlining from your own notes | Lower | Keep the argument yours; rearrange freely |
| Summarising papers you have read | Medium | Verify against the original before citing anything |
| Generating references or statistics | High | Verify every item in the original source; remove what you cannot verify |
| Drafting interpretation or conclusions | High | Write interpretation yourself; AI text here is rarely defensible |
| Entering unpublished or sensitive material | High | Do not use public tools; use institution-approved tools only |
These authoritative external resources explain the standards and guidance mentioned in this article:
Is it allowed to use AI tools when writing a research paper?
Most journals allow AI tools for language improvement and similar assistance, provided the use is disclosed where required and the authors take full responsibility for the content. Always check your target journal's current policy, because expectations differ and change.
Can I list an AI tool as a co-author?
No. Major publication guidance states that AI tools cannot be authors, because authorship requires accountability that a tool cannot provide. Human authors remain responsible for every word in the manuscript.
Do I need to disclose AI use for grammar checking?
It depends on the journal. Many journals ask for disclosure of any AI assistance, including editing, while others focus on substantive generation. When in doubt, disclose: an unnecessary disclosure is harmless, while a hidden one can be treated as misconduct.
How do I check AI-generated references?
Look each reference up in the original source — the journal's site, a database, or the publisher's page. If you cannot verify a reference, remove it, however well it fits your argument. Fabricated citations are a common and easily detected problem.
Is AI-generated text plagiarism?
It can be, in two ways: by reproducing wording from training material, or by paraphrasing so closely that the result is effectively copied. Run plagiarism checks as usual, and rewrite any passage that depends on wording you did not originate.
Can I paste a manuscript under review into an AI tool?
No, not into a public tool. Manuscripts under review are confidential, and material entered into public tools may be stored or used beyond your control. Use only institution-approved tools with appropriate protections for sensitive material.