HumanizersFor students
Rewording a Sentence ChatGPT Wrote, Uncited, Is Plagiarism Under Codes Like USC's and CUNY's
Rephrasing turns on who wrote the sentence first. A student who drafts a paragraph and asks ChatGPT to tighten it is editing. MLA asks for that to be acknowledged [3], and each syllabus permits or forbids it on its own terms [4]. Asking ChatGPT for the paragraph and rewording it without citing it is something else: USC's and CUNY's codes put paraphrase inside their plagiarism clauses [1][2], and a student's own institution may word it differently. A detector reading on the result doesn't say who wrote it.
Olive Staff & OliveUpdated
Where Does Rephrasing Stop Being Editing?
At the origin of the sentence. If a student drafted the words and then asked ChatGPT to tighten them, the result is an edited draft. If ChatGPT produced them first and the student reworded them afterwards, it's machine writing with a new surface. The University of North Carolina writing center puts the test plainly: an AI tool should help a writer think and write rather than think or write for them 4.
The amount of rewording doesn't decide it.
Two paragraphs can differ by one word and land on opposite sides of a misconduct code, one a student's argument in slightly better clothes, the other a chatbot's argument in a disguise the student picked. Real papers are messier. Three paragraphs get written on a Tuesday afternoon, a fourth is requested from ChatGPT at 11pm, two of its sentences survive and the rest get reworded. By Thursday nobody can say which was which.
That paper has a real authorship answer. There's no record of it, though.
Plagiarism and breaking a course rule are separate questions, and the codes answer them differently: is using ChatGPT plagiarism.
What Do University Codes Say About Paraphrased AI Text?
Two codes cited here name paraphrase directly. The University of Southern California's academic integrity office says material created by a generative tool and represented as a student's work is plagiarism whether it was copied verbatim, near-verbatim or paraphrased 1. CUNY gets to the same place from the other side. Its academic integrity policy lists use of AI-generated content, even when paraphrased, without citing the AI as the source 2.
MLA's style guidance asks for less than a misconduct charge and more than silence. MLA wants every functional use of an AI tool acknowledged, editing prose included, and a note or a line in the text will do 3. It does not say a rewording pass is forbidden. It says the pass gets written down 3.
Whether a course allows the pass at all is often the syllabus's call. The same North Carolina page asks a student to consult each syllabus for its requirements on AI tools 4. Two courses in one department can draw the line in different places in the same term.
A disclosure sentence written before the deadline costs a line, and it puts on record what no detector reading can show: how to write an AI disclosure statement.
Does a Rephrased Paragraph Get Flagged?
Sometimes, and not reliably in either direction. A rephrased paragraph is mixed writing by construction, and detectors handle mixed writing badly. A 2024 benchmark was built from exactly that material: human text revised by a model, and machine text revised by a person. Existing detectors struggled with it, particularly with subtle modifications and changes of style 6.
Bloomberg Businessweek tried something closer to what students do. Asking Grammarly to improve an essay, or to make it sound academic, turned writing that had passed as 100% human-written into 100% AI-written. Ordinary spelling and grammar fixes barely moved the reading 5.
The essay's authorship didn't change between the two readings.
Misses go the other way too. In Pangram's own technical study, as reported by Nature, when Pangram Labs tested consumer AI used to substantially modify human-written student essays, its model still called 41% of the results fully human 7. No reading describes conduct. What one justifies is opening the draft history and asking about the argument, because a flag is never a finding. A dedicated rewriter is its own case: a QuillBot rewrite and what a detector reads.
Rewrite From the Student's Own Sentence
Draft the paragraph badly first, then ask for specific changes to it. A request to shorten the second sentence or cut the hedges works on writing that already exists, and the file keeps the version from before the paste. Ask for a whole paragraph instead and the student gets something to own or disown.
| Into the tool | Back out | What USC's and CUNY's clauses call it |
|---|---|---|
| A paragraph the student wrote | Tidier | An edited draft, disclosable 3 |
| A prompt about the assignment | A paragraph nobody drafted | Plagiarism if submitted as the student's own without citing the AI 12 |
| AI text put through a rewriter | Same argument, disguised | Still AI content; paraphrase does not remove the need to cite 12 |
Keeping the earlier version is most of the defense.
Google Docs keeps a version history that anyone with edit access can browse 9. A paragraph pasted into the file finished arrives with no drafting history behind it. Rewriting from notes changes more than vocabulary, and it keeps the argument the student's own: humanizing AI text by hand.
One finding cuts against this advice. The 2026 Notre Dame study, run on published research abstracts rather than student papers, found lightly AI-edited text flagged far more often than humanizer output, and its authors describe that as a higher sanction risk for honest AI editing 8. That's a case for fixing the policy. Dedicated rewriters raise a question of their own: is QuillBot cheating.
Check the Draft, Revise the Flagged Sentences, Check Again
For a paragraph the student wrote and then had a tool reword: check the draft, and rewrite the sentences that come back machine-edited from the notes. The loop is about the student's own prose reading as their own. It won't turn a paragraph ChatGPT wrote first into the student's work; that paragraph needs a citation or a fresh draft.
A writing tool is in private preview that reads a draft with the same detector, colors each sentence by what the detector reads it as, keeps quotations, numbers, citations and anything the writer marks as their own untouched, and helps rewrite the rest in the writer's own words; it never tells a person that their own writing is not theirs. What a student can run today is the check, revise and re-check loop: paste up to 1,200 words and get a reading, one paper a day, with no account. Human's detector scores the text in memory and de-identifies any personal information for additional training.
Human, at human.olive.is, reports how much of a document reads as machine-written or machine-edited, sentence by sentence, and gives a document verdict of Human, Mixed or AI. A sentence it marks is one that reads that way, which is not a record of what a tool touched. College-level academic writing is what it's tuned for. A single reworded sentence checked on its own gets nothing back, since under 50 words the detector returns no reading at all.
Its three operating points are Human's own in-house measurements rather than an independent audit: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%.
On its own held-out sets, measured September 2026, Human's detector recorded 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. On the machine side it recorded 0 misses in 481 wholly machine-written documents; the ceiling on that is 0.62%. Neither set is reworded or mixed text, the kind this article is about; carrying these rates over to a rephrased paragraph is an assumption, and edited text is harder for a detector to read. Those are in-house numbers, measured by the company that makes the detector. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The reading is a diagnostic, and the writing is the point.
Common questions
Is asking ChatGPT to reword one sentence cheating?
Not by default, if the sentence was the student's own to begin with. That's an editing pass. MLA guidance treats it as a functional use of the tool, to be acknowledged rather than treated as misconduct 3. The course policy is what can change the answer. The North Carolina writing center sends a student to each syllabus, because instructors' views may vary from one class to the next 4. A sentence ChatGPT wrote in the first place is a different case, however the wording changes afterwards.
Does rewording AI text remove the need to cite it?
No. CUNY's academic integrity policy lists use of AI-generated content, in whole or in part and even when paraphrased, without citing the AI as the source 2. USC's integrity office is built the same way: it covers material created by a generative tool and represented as a student's work, whether paraphrased, copied verbatim or near-verbatim 1. Paraphrase is exactly what both clauses were written to reach. Rewording adds work and doesn't change what the code calls the submission.
Can a detector tell a rephrased draft from a rewritten one?
Not reliably, and the published measurements show it from both sides. A 2024 mixed-text benchmark found existing detectors struggled with it, particularly with subtle modifications and changes of style 6. Nature reported a figure from Pangram's own technical study: in tests of consumer AI used to substantially modify student essays, Pangram's model still called 41% of the results fully human 7. A reading describes text. It settles nothing about who did what, so the draft history and a conversation about the argument have to do the work a percentage can't.
What counts as evidence that the student wrote the original sentence?
A version history, mostly. A document drafted over hours carries timestamped revisions. A paragraph pasted in finished carries one. Notes, outlines, marked-up sources and a messy earlier draft point the same way. None of it matters until someone asks, and that's the argument for writing in a file that keeps its own history, before any flag exists, instead of composing somewhere that leaves no drafts in the file.
Is rephrasing in ChatGPT different from running text through QuillBot?
USC's clause does not separate them: it covers AI-created material represented as the student's work, whether paraphrased or copied 1, and does not ask what did the paraphrasing, though other codes may differ. A general chat tool rewords whatever it's handed, and a dedicated rewriter does the same job with a narrower menu. Human's own in-house finding, on its own detector and its own test sets, is that whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. That is not an independent audit and says nothing about other detectors.
Does disclosing a rewriting pass make an accusation more likely?
No study cited here measured disclosure itself. The 2026 Notre Dame paper tested published research abstracts, not student submissions. It found lightly AI-edited text flagged far more often than humanizer output, and its authors summarize that as honest AI editing carrying a higher sanction risk than humanizer-assisted evasion 8. Whether that carries over to how an institution treats a student's disclosed edit is an assumption. A disclosure line written before submission is still cheap evidence to have if a reading comes back mixed later.
References
- 1.Academic Integrity & Generative AI University of Southern California, Office of Academic Integrity, 2026. academicintegrity.usc.eduThe statement that work authored by another, including material created by ChatGPT and other generative AI tools, but represented as the student's work is plagiarism whether paraphrased or copied verbatim or in near-verbatim form.
- 2.Academic Integrity Policy The City University of New York, 2022. cuny.eduThe clause naming unauthorized use of AI-generated content, or use of AI-generated content whether in whole or in part, even when paraphrased, without citing the AI as the source.
- 3.How Do I Cite Generative AI in MLA Style? MLA Style Center, 2023. style.mla.orgThe instruction to acknowledge all functional uses of the tool, such as editing prose or translating words, in a note, the text, or another suitable location.
- 4.Generative AI in Academic Writing The Writing Center, University of North Carolina at Chapel Hill, 2023. writingcenter.unc.eduThe test that AI tools should help a writer think and write rather than think or write for them, and the instruction to consult each syllabus for guidance or requirements related to the use of AI tools.
- 5.AI Detectors Falsely Accuse Students of Cheating—With Big Consequences Bloomberg Businessweek, 2024. cs.utexas.eduThe finding that using Grammarly to improve an essay or make it sound academic turned work that passed as 100% human-written into 100% AI-written, while non-generative corrections barely moved the reading. The url is a course-hosted copy of the Bloomberg Businessweek article; the passage does not name the detector used.
- 6.LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected? arXiv (NAACL 2024 Findings), 2024. arxiv.orgThe dataset built from mixed writing, where a person revised machine text or a model revised human text, and the finding that existing detectors struggle to identify it, particularly with subtle modifications and style adaptability.
- 7.AI-detection tools have made huge leaps forward — how good are they? Nature (Naddaf and Van Noorden), 2026. nature.comNature's report of a figure from Pangram's own technical study: when Pangram Labs tested consumer AI used to substantially modify human-written student essays, its model still labeled the results fully human 41% of the time.
- 8.Why AI Detection Fails for Academic Integrity Karr, Khvatskii, Hua and Chawla, University of Notre Dame (arXiv), 2026. arxiv.orgA study of detector flag rates on published research abstracts, whose abstract states that honest AI editing results in a higher sanction risk than humanizer-assisted evasion.
- 9.Find what's changed in a file Google Docs Editors Help, 2026. support.google.comThe version history of a Google Docs file, which a person needs edit permission to browse.
9 sources, numbered by first appearance.
General guidance for teachers, administrators and students. What holds at one institution, on one assignment, may not transfer to another.
Human checks whether a piece of writing can be verified as written by a person, and shows the sentences behind the result. It does not report a probability that a person used AI, it does not check for plagiarism, and no number it produces stands for a student's honesty. A flag should be reason to look closer and not verified proof of cheating.