How much editing makes AI text your own?

The short answer: editing changes the words, not the authorship

There is no percentage. No amount of rewording turns a model's paragraph into your paragraph, because authorship in academic work is not measured in edited words. It is measured in contribution and accountability.

That sounds like a dodge. It is not, and the formal standards are unusually specific about it.

The International Committee of Medical Journal Editors publishes the most widely used authorship criteria in academic publishing. It lists four, and all four must be met: substantial contributions to the conception or design of the work, or the acquisition, analysis, or interpretation of data; drafting the work or reviewing it critically for important intellectual content; final approval of the version to be published; and agreement to be accountable for all aspects of the work.

One caveat, because you will see this standard inflated. The ICMJE stopped maintaining its list of journals claiming to follow its Recommendations in April 2025, citing its inability to verify the list's accuracy and the fact that many listed journals did not actually adhere. No honest count of adopting journals exists. The criteria themselves are unchanged, and it is the criteria that matter here.

Read the second one again. Drafting the work is a criterion, but on its own it is not sufficient. The ICMJE's guidance on non-author contributors says so directly: writing assistance, technical editing, language editing and proofreading are activities that alone do not qualify a contributor for authorship. Those people are acknowledged, not listed as authors. The logic runs the other way too. If you did not supply the ideas, editing the sentences will not promote you to author.

Whether using a rewriting tool at all is defensible is a separate question, and one we answered in our piece on whether an AI humanizer counts as cheating. This article is about the line itself.

What "substantial contribution" means, and why it is really about accountability

Every serious authorship standard converges on the same idea, and the idea is not originality. It is responsibility.

The ICMJE's fourth criterion asks authors to agree to be "accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved." Nature Portfolio's editorial policy on AI states the same thing in one line: "Accountability for scholarly content, evaluation, and editorial decisions cannot be delegated to AI systems."

That is also why the ICMJE says chatbots "should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work." A model cannot be asked what it meant. It cannot be called into a viva. It cannot answer for a fabricated citation.

Someone has to be able to answer, and the practical version is concrete. Can you defend every claim, explain why the argument is ordered the way it is, and produce the source behind every number? If yes, you are behaving like an author. If three follow-up questions from your marker would expose you, the words being yours will not help.

Why swapping words fails: the patchwriting problem, named in 1993

Writing teachers had a name for this problem three decades before ChatGPT.

In "A Plagiarism Pentimento," published in the Journal of Teaching Writing in 1993, Rebecca Moore Howard named the practice "patchwriting": "copying from a source text and then deleting some words, altering grammatical structures, or plugging in one-for-one synonym substitutes." Howard's argument was sympathetic: she saw patchwriting as a learning stage, an outsider's attempt to speak a new academic language. She was also blunt about where it stands: "Valued members of the academic community do not patchwrite."

It is common. The Citation Project, a multi-institution study led by Sandra Jamieson and Rebecca Moore Howard, analysed 174 researched papers by first-year students at sixteen US colleges and universities. It found at least one instance of patchwriting in 52% of the papers, and that in 94% of citations the students were working from only one or two sentences of their source.

Now swap the source. Deleting words, altering grammar and substituting synonyms is precisely what a paraphrasing pass over a model's draft does. The mechanism is identical, which is why running AI text through a rewriter does not clear the authorship bar. Whether it still reads as machine-written is a separate question, covered in our piece on paraphrasing tools and Turnitin.

One fairness note. Jamieson and Howard argued in a 2019 paper in the International Journal of Technologies in Higher Education that textual errors such as patchwriting should be treated as bad writing to be remedied by teaching, not as breaches of integrity to be punished. Many universities do not draw that distinction. Carnegie Mellon's Eberly Center publishes sample course policies, one of which states plainly that "passing off any AI generated content as your own (e.g., cutting and pasting content into written assignments, or paraphrasing AI content) constitutes a violation of CMU's academic integrity policy."

The spectrum from generated to genuinely yours, in five levels

The most useful published map of this territory is the AI Assessment Scale, developed by Mike Perkins, Leon Furze, Jasper Roe and Jason MacVaugh in the Journal of University Teaching and Learning Practice in 2024. It sets out five levels of permitted AI use. It was written for teachers, but read it as a student: it shows what your institution is choosing between.

Level 1, No AI: "the assessment is completed entirely without AI assistance." Level 2, AI-Assisted Idea Generation and Structuring: AI for brainstorming and outlining, but "no AI content is allowed in the final submission." Level 3, AI-Assisted Editing: "AI can be used to make improvements to the clarity or quality of student created work to improve the final output, but no new content can be created using AI," with your original AI-free work supplied in an appendix. Level 4, AI Task Completion with Human Evaluation: the model completes specified parts and "any AI created content must be cited." Level 5, Full AI. A 2025 revision by Perkins, Roe and Furze renames the middle levels and adds an "AI exploration" level.

Look at Level 3 closely, because it is where most anxious students think they are standing. The condition is not how much was edited. It is that no new content came from the model, and that you can still produce the version you wrote before it was involved. That is the real threshold, in a framework its authors report has been implemented in hundreds of institutions worldwide and translated into 30 languages.

Which direction the editing runs is the whole question

Two students both say "I edited it." One wrote a draft and asked a model to tighten the prose. The other asked a model for a draft and tightened the prose themselves. The same number of words changed. Completely different situations.

Nature Portfolio's editorial policy on AI is the clearest public statement of why. It sorts AI use by risk rather than by tool. Green, permitted: "AI use that supports expression, organisation, or efficiency without influencing scientific, scholarly or evaluative judgement," where "AI use does not introduce new intellectual content." Amber, exercise caution: use where "AI does introduce new intellectual content and requires verification and oversight," with "extensive copy editing or writing support" given as an example. Red, not permitted: "generating hypotheses, analyses or conclusions and presenting them as human-derived."

The dividing line is new intellectual content, and it does not move when you rewrite the sentences afterwards. If the model supplied a claim, a structure, an example or an interpretation you did not have, that content is in your document no matter how the sentences read now.

This is also why "I only changed the wording" is the wrong defence. It concedes the point. It says the ideas came from somewhere else.

A five-question test you can run on your draft tonight

None of the above helps at 1am with a draft open, so here is the concrete version. Run all five. A single "no" is your answer.

1. The closed-laptop test. Close the draft. On a blank page, write the argument in five bullet points from memory. If you cannot reconstruct your own thesis and its supporting moves without looking, the reasoning is not yours yet.

2. The source test. Pick three factual claims at random. Open the actual source for each one. If a claim traces back only to the model, it is not evidence. It is a sentence that sounds like evidence.

3. The structure test. Ask yourself why section three follows section two. If the honest answer is "that is the order it came out in," the architecture of the argument belongs to the model.

4. The substitution test. Take one paragraph and compare it against what the model produced. If the differences are synonyms, deletions and rearranged clauses, that is Howard's patchwriting, not rewriting.

5. The disclosure test. Write one sentence describing exactly what you did, in the form Carnegie Mellon's sample policy suggests: "I generated this work through ChatGPT and edited the content for accuracy." Now picture it at the top of your submission, read by your marker. If that changes how you feel, it is telling you something your word count is not.

Five yeses means you are on defensible ground, and you should still check your course's disclosure rules. A single no means there is work left, and the work is not editing.

Worked example: three versions of the same paragraph

Start with a model output on a familiar topic. The specifics are illustrative; yours would come from sources you actually read.

Model draft: "The Industrial Revolution fundamentally transformed British society, ushering in unprecedented urbanisation and reshaping the relationship between labour and capital in ways that continue to resonate today."

Version A, patchwritten: "The Industrial Revolution completely changed society in Britain, bringing about unmatched urban growth and altering the connection between workers and capital in ways still felt now." Every content word has a substitute. The claim, the ordering and the framing are untouched. Nothing was added and nothing was checked. This is Howard's 1993 definition almost line by line, and it is what most students produce when trying to make AI text "theirs."

Version B, written from understanding: "Britain's cities grew faster than their housing and sanitation could absorb, and that is the part the word 'urbanisation' hides. The contemporaries I read describe a housing and public health emergency first and a change in economic relations second. Calling it a transformation in the relationship between labour and capital is accurate but too smooth."

Version B took longer. What actually changed: the writer went to sources, added a specific the model did not supply, and pushed back on the model's framing. That pushback is the contribution. It is also, not coincidentally, better writing. Version B is the only one of the three you could defend under questioning.

Disclosure rules genuinely differ by course, and that is not a loophole

There is no single answer to "do I have to declare it," and anyone who offers you one has not looked.

Harvard's Faculty of Arts and Sciences publishes three sample AI policies for instructors. One forbids "the use of ChatGPT or any other generative artificial intelligence (AI) tools at all stages of the work process, including preliminary ones." One encourages AI "for all assignments and assessments" as long as the use is "appropriately acknowledged and cited." One sits in between. All three close with the same warning: "different classes at Harvard could implement different AI policies, and it is the student's responsibility to conform to expectations for each course."

Carnegie Mellon's Eberly Center publishes six examples across the same range, from a ban that explicitly includes brainstorming to a policy permitting verbatim AI content with quotation marks and a citation.

So the thing to find is not a general rule. It is your specific course's rule, usually on the syllabus or course page. Ask if it is not there. That email costs you nothing and it is a complete defence.

When disclosure is required, the style bodies have worked out the mechanics. The MLA Style Center says to cite a generative AI tool "whenever you paraphrase, quote, or incorporate into your own work any content" it created, and separately to "acknowledge all functional uses of the tool (like editing your prose or translating words)." Those are two obligations, and paraphrasing does not remove the first. APA's guidance asks you to describe how you used the tool in your method section or introduction, give the prompt you used, and put long responses in an appendix, because a model produces a different response in every session and your reader cannot reproduce it.

When the honest answer is "this is not your work yet"

Sometimes the test comes back negative and you knew it before you finished reading the list. The model wrote the argument. You changed the words. It is due tomorrow.

Here it is plainly: rewording it further does not fix it, because the wording was never the problem. Text does not become yours by being edited enough times.

What does fix it is cheaper than it sounds. Pull the draft's claims into a bullet list and throw the prose away. Check each against a real source. Some survive, some turn out vague, and one or two turn out to be wrong, which is the most useful thing that will happen to you all evening. Then write the paragraphs from your bullets and your sources, in your order. The argument usually changes, because you now disagree with parts of it, and that disagreement is your contribution.

If you have done that work, the reasoning in the document is now yours, and using a rewriting tool to clean up phrasing is an editing pass over your own writing. That is the Level 3 situation described above, and note that Level 3 also asks you to keep the AI-free original for an appendix. If your course does not permit it, no tool changes that, and nobody selling you one can tell you otherwise.

If you wrote it yourself and you are still worried about being flagged

Here is the uncomfortable flip side. Plenty of students who wrote every word get flagged anyway. Detection is probabilistic and false positives are real. A 2023 study in Patterns by Weixin Liang and colleagues at Stanford tested several widely used GPT detectors and found they consistently misclassified non-native English writing samples as AI-generated, while native samples were identified correctly. A flag is not a finding of fact, and it is not an accusation. What a flag actually means, and how to demonstrate authorship is worth reading before you panic.

Two practical moves. First, run your own draft through a free AI detector before you submit, so you find out in private rather than in a meeting. Treat the number as a prompt to look at your writing, not as a verdict on it.

Second, if the score is high on work you genuinely wrote, look at what made it read that way. A scan for stock AI vocabulary will show you the specific words to reconsider. Rewriting those passages in your own idiom improves the essay regardless of what any detector says.

And keep your evidence as you go, not afterwards. Version history, dated notes, annotated reading, earlier drafts. Authorship is easiest to demonstrate with process, and process is the one thing a model cannot hand you.