Which Writing Tools Count as "AI Assistance" You Have to Disclose? A Checklist

The narrow endpoints are clear in the publisher documents below: Elsevier says basic spelling, punctuation, and grammar checks do not need a declaration, while IEEE requires disclosure of model-generated content used in an article. The awkward middle — rephrasing, translating, outlining, auto-generating references — is where the published documents stop agreeing with each other, and that is exactly where the tools people ask about live. Below is a row per tool with the rule that decides it. None of it tells you what your own institution requires, and nothing here is a judgment about whether any particular use is acceptable.

Start with the sentence that dissolves most of the confusion. Nature Portfolio's current AI framework says it outright, at the top of the table it uses to sort everything else:

"Use of AI is governed according to risk level, not by tool type."

A product-name list cannot classify the cases in this article. Grammarly's own help separates spelling and grammar support from paragraph-level rewrites; a chat assistant can fix a comma or write your discussion section. The rules quoted below classify what happened to the text, which means the unit you have to classify is the action you took, not the icon you clicked.

Each row names a published document. Where a body has been quoted, the words are theirs. The fourth column is the part worth reading twice, because that is where a row flips.

Read the rows as "this is what these documents say", not as "this is your answer". Publishers differ from each other, universities differ from publishers, and a single department can be stricter than the university it sits in. We compared six publishers' declaration requirements separately in publisher AI disclosure policies, compared.

When a tool is not on the list, or the interface changed since you last looked, these get you to the right row.

1. Did text exist afterwards that did not exist before? Correcting a comma leaves your sentence. A suggestion you accept with one click can be a sentence the model composed. Same subscription, different rows.
2. Does your rule cover assistance, or only generation? A rule written about "AI-generated content" and a rule written about "AI assistance" ask for different declarations from the same person, and one policy can use both phrases in different paragraphs.
3. Is the rule about the text, or about who decided? Auckland's working-methods section requires substantive amendments to "be indicated using comment tools rather than making direct changes", so that the candidate "actively considers and makes decisions about all substantive suggestions". A tool can break that rule by the shape of what it hands back, without a single word being objectionable.

The third question matters because a tool can be within a permitted-changes list and still fall foul of the same document three paragraphs later.

The clearest single sentence on editing comes from IEEE, and it is worth reading with an eye on the verbs, because they change halfway through:

"The use of content generated by artificial intelligence (AI) in an article (including but not limited to text, figures, images, and code) shall be disclosed in the acknowledgments section of any article submitted to an IEEE publication. The AI system used shall be identified, and specific sections of the article that use AI-generated content shall be identified and accompanied by a brief explanation regarding the level at which the AI system was used to generate the content. The use of AI systems for editing and grammar enhancement is common practice and, as such, is generally outside the intent of the above policy. In this case, disclosure as noted above is recommended."

Three shall, then one recommended. Generated content is mandatory to disclose and must be pinned to a section; editing and grammar enhancement are carved out and downgraded to advice. We are a rewriting tool and it would suit us to stop the quote one sentence early, so: the carve-out is for editing and grammar enhancement, and IEEE does not define where those end.

Elsevier draws the same boundary at a different place and, unusually, says what pushes you across it:

"Basic checks of grammar, spelling and punctuation do not need a declaration statement. However, when an AI tool makes substantive changes to sentence structure or organization of a part of the text, this should be disclosed."

Nature Portfolio divides the middle band differently. Its risk framework puts "using an LLM to polish or refine the language" in green, alongside suggestions about manuscript-section structure or formatting and work to translate, structure, or clarify reviewer comments. It puts "extensive copy editing or writing support" in amber, which it describes as permitted with human oversight and disclosure. A language polish and a heavy copy edit therefore land in different bands of the same table; neither the table nor the page gives a numerical threshold between them.

MLA has moved on this too, and the change is visible if you open both posts. Its 2023 guidance on citing generative AI told writers to "acknowledge all functional uses of the tool (like editing your prose or translating words) in a note, your text, or another suitable location". That post now carries a banner pointing at an updated August 2025 version, which drops the list and sends readers to a separate post on acknowledgement — where the standard is to "acknowledge substantive uses of AI", with the definition of substantive handed off to a cited journal article rather than fixed by MLA itself.

The honest summary is narrower. Elsevier exempts basic checks but requires disclosure when AI makes substantive structural changes; IEEE mandates disclosure for generated content but only recommends it for editing and grammar enhancement; Nature Portfolio separates language polishing from extensive editing or writing support. Auckland's thesis rules add a broader requirement to declare covered GenAI use. These are scoped rules, not a cross-institution consensus, and your institution may draw the line somewhere else.

People collapse two requirements into one and then answer the wrong one. Disclosure tells a reader that a tool was involved. Citation points at a specific piece of content and lets a reader locate it. Some manuals only tell you how to do the second.

OSCOLA's rule 3.7.13, "Generative artificial intelligence", is a citation rule: cite the AI as the author, put the instructions you gave it in quotation marks, name the organisation that developed it, and give the generation date. Its own worked example is `ChatGPT, response to 'Explain how artificial intelligence works', OpenAI (16 July 2023).`

MLA keeps the two apart explicitly. Its acknowledgement post says that "acknowledgment of the use of AI is slightly different from, and more comprehensive than, citation", because "generative AI output is not a source in the traditional sense" — it assembles text by prediction rather than reporting something you could go and read. A citation covers the passage you took; when the use is substantive, MLA advises acknowledging that use as well.

That distinction is what makes the reference-manager row above less obvious than it looks. Formatting a bibliography is not a disclosable use in Elsevier's wording, but asking a tool to select or generate the references is, and the two now sit close together in the same software. Whatever you do there, the accuracy of every reference is yours: the ISTIC-led boundary guideline is blunt that authors must verify each suggested reference themselves, because generated ones can be fabricated or out of date.

Nothing above is about detectors, and mixing the two questions is how people end up doing the wrong thing at the wrong time. A disclosure obligation is a rule your institution or publisher wrote about your process. A detector output is a measurement of the finished text. You can be entirely inside the rules and still be holding a report with a number on it, and the reverse is also possible.

The rulebooks quoted here really do stay in their lane, and that is checkable rather than a matter of opinion. IEEE's editorial style manual came to 141,011 characters when all 67 pages were extracted with `pdfplumber`, joined with a newline and had runs of whitespace collapsed to a single space. Searching then has to happen on a second variant with all whitespace stripped out, 131,576 characters, because that PDF's text layer drops spaces between words — on the 141,011-character version `artificial intelligence` returns 0, and that zero is false; strip the spaces and it returns 2. On the stripped text, `Turnitin` appears 0 times, `plagiar` 0, `AI detection` 0, `detection` 0, `integrity` 0, `misconduct` 0. The instrument was alive on the same run: those two `artificial intelligence` hits are the disclosure clause above and the glossary entry `GenAI generative artificial intelligence`, and `detector` appears twice, both times in the same abbreviations glossary (`maximum-likelihood sequence detector`, `resistance temperature detectors`). OSCOLA's fifth edition is 85 pages and, under the same treatment with `pypdf`, 174,071 characters. There `detect` returns 0, `turnitin` 0, `plagiar` 0, `integrity` 0, while `artificial intelligence` returns 11 and `generative` 7 — and those six counts are the same whether you collapse the whitespace or strip it. Both PDFs were downloaded on 28 August 2026.

Read that both ways, because it cuts against us as well. It means these two manuals are not telling anyone what a detector will do with an edited paragraph — but it equally means they are not blessing anything either. The absence of the word `detection` in a style manual is not permission.

For the detector half of the question, the vendor has answered separately and at some length about grammar tools specifically, including the qualifier that limits the answer: does Turnitin detect Grammarly as AI writing. And if you already used several of these tools on one draft and got a number back, I polished my essay with Notion AI and QuillBot and now it is flagged is the version of this written from that end.

Five steps, in the order that saves the most rework.

1. Find the governing document, in writing. Your academic integrity policy or your target journal's author instructions. If you are writing a thesis, look for a third-party editing or proofreading document as well — it may be filed separately from the AI guidance and can impose a different rule. We took three of those apart in is an AI rewriting tool covered by your university's third-party editing rules.
2. Check which noun the rule uses. "AI-generated content" and "AI assistance" are different scopes, and a document can contain both.
3. Classify by action, one action at a time. One chat session can hold four different uses. Then record both the tool and what it did: Elsevier's declaration template asks for the tool or service name, the reason it was used, and the author's review of the resulting content.
4. Keep a use log while you work, not afterwards. Tool and version, date, what it did, which section, what you checked. Reconstructing this at submission is where the inaccuracies come from.
5. Ask the editorial office or your supervisor when a row is genuinely ambiguous, and keep the answer. A short written reply from the people who administer the rule outranks every paragraph on this page.

If a use turns out to be prohibited rather than disclosable, disclosing it does not convert it into a permitted one. That sounds obvious written down and it is not obvious at 2am.

A commercial disclosure before applying the checklist: HumanPen is our product, so this paragraph describes its current workflow rather than offering an independent recommendation. You upload the document and mark the passages you want rewritten, or provide a detection report you already have so the relevant passages can be located; if a selection stops mid-paragraph, the app expands it to the whole paragraph and asks you to confirm the scope before processing. Processing is limited to that confirmed range, and the result is an editable document designed to preserve your citations, numbers, tables and layout; where the result qualifies, you can carry on lowering the AI score at no extra charge.

On the checklist above, that is the "rewrites a paragraph you wrote" row. Under Elsevier's wording, whether it crosses the substantive change threshold depends on what changed in that use. Under Auckland's guideline for theses and dissertations, third parties or GenAI must not rewrite any part of the covered thesis text; scope controls do not override that clause. We would rather write both limits here than have you find either one after submission.

HumanPen cannot classify your particular edit for you. Whether one changed paragraph in your discussion section counts as "substantive" under your journal's wording, or as "editing and grammar enhancement" under IEEE's, is a judgment your editor or your institution makes. Our ethics page keeps that boundary with the people who administer the rule: processing a manuscript does not erase any disclosure requirement that already applied, because rewording changes the wording rather than the history of how it was written. The full reasoning is in why rewriting does not erase a disclosure duty.

My university's policy just says "AI tools" with no list. What do I do? Classify by action and write down which rows you landed on and why. A short note saying "I used X for Y in section Z, and here is what I checked" answers the question a policy like that is actually asking, and it is a much better artefact than a guess about whether a product name counts.

Is a spell checker really exempt everywhere? Elsevier says integrated spelling and grammar checkers are exempt from its policy. That is not a universal rule. If your institution's policy says any tool, it means any tool, and it is entitled to.

I ran my own English through a translator. Is that AI assistance? At Elsevier, yes, and it says so directly: you remain fully accountable for the translated text and should disclose the use for translations included in your manuscript. Whether it also affects a detector score is an unrelated question with its own material, in does machine translation get flagged as AI.

If I disclose, does that make anything allowed? No. Permission and disclosure are separate requirements, and Auckland is the clean example: acknowledgement of GenAI use is required there, rewriting is prohibited there, and satisfying the first does not satisfy the second.

Does disclosing hurt me? That depends on the rule and context. Nature Portfolio says transparent declaration "removes ambiguity about how AI might have been used"; that sentence does not predict how your institution will treat a particular declaration. If you think it may be read differently in your context, ask a supervisor before submission, not after.

Do these rules apply to coursework, or only to journal articles and theses? The publisher policies apply to manuscripts, and Auckland's document is written for theses and dissertations. Coursework may be governed by a different document at the same institution. Read the one that covers what you are submitting.