A Humanizer Rewrites Wording; It Does Not Settle Plagiarism

Different things, on different evidence. A similarity or originality report matches the submitted words against other documents: the Google Classroom help page, current in 2026, describes comparing files against web pages and books on the internet, plus earlier student submissions where a school turns that on, and nowhere describes AI writing detection ². The other checker reads the prose and estimates how much of it looks machine-produced.

One asks where the words came from, the other asks what wrote them. Neither was built to decide the case: Turnitin wrote in 2023 that it does not make a determination of misconduct even in the space of text similarity ³. That is the older and better understood of the two instruments, and the newer one has not earned more deference.

A student who meets both numbers in the same week can read them as a single verdict. A reading of each, side by side, is here: what a similarity score and an AI score each measure.

No, and two large policies say so outright. Pangram Labs, in a 2025 explainer, defines a humanizer as a tool that rewrites machine-written text to evade AI detectors ¹. The University of Southern California's integrity office counts material created by a generative tool and presented as a student's own work as plagiarism whether it was copied verbatim, near-verbatim or paraphrased ⁴. A rewrite is a paraphrase.

A different route across the country arrives at the same place. The City University of New York's 2024 academic integrity policy names the use of AI-generated content, in whole or in part, even when paraphrased, without citing the AI as the source ⁵. Two large systems, one on each coast, both put paraphrase inside the plagiarism clause.

What a rewrite moves is a string-match number. Reworded copy matches fewer strings, so the similarity figure falls and the missing citation stays missing. The rule was never about the wording.

Where editing a student's own sentence stops and rewording a machine's begins is worked through in whether rephrasing ChatGPT output counts as cheating.

Almost nothing. A zero on a similarity report means the matching engine found nothing to match. A paper whose words appear nowhere in that index comes back that way. So does a paragraph a chatbot produced this morning, because the report compares against web pages, books and, where a school turns it on, earlier student submissions ². Low AI readings after a rewrite say just as little about where a draft came from.

Someone has measured that second point. Researchers at the University of Notre Dame published the count in 2026, after running machine-written rewrites of 642 abstracts, 25 to 500 words each, through the humanizer Undetectable AI. Fewer than 4 percent stayed flagged by the two detectors tested, Pangram 3.2 and GPTZero, a false negative rate above 96 percent on that set ⁶. The reading fell. The origin of the text did not change.

That reading gets over-read in both directions. Turnitin's 2023 guidance is that it does not determine misconduct ³, and an integrity office that treats a percentage as the case has skipped the part where a person reads the draft, the sources and the history behind them.

On the tool pages read for this article, the promise is about a detector score, not about a citation. The BypassGPT homepage tells students the tool "can consistently help you submit essays and assignments without breaching any academic guidelines" ⁷. That is a claim about institutional rules, and no integrity policy read for this article supports it. A tool page cannot know what a course policy says.

Some of what a humanizer does to a page is visible to a marker with no tool at all. The same 2025 Pangram Labs explainer documents one technique: adding nonsensical phrases, in the hope a detector decides gibberish is unlikely to have come from a machine ¹. Filler reads as filler, and the grade arrives from a person. Marketing carries no standing in an integrity code.

What a rewriting pass leaves behind, sentence by sentence, is its own question: what a humanizer does to writing.

Attribution settles the citation half of the plagiarism question. The 2024 CUNY policy lists two things under plagiarism: unauthorized use of AI-generated content, and use of it, in whole or in part, even when paraphrased, without citing the AI as the source ⁵. So at CUNY a citation settles the second and not the first, and whether the use was allowed is the syllabus's question.

Disclosure is not the same as permission. Where a course allows AI assistance with acknowledgement, the citation formats are a separate walkthrough: how to cite ChatGPT in APA and MLA.

Checking a finished draft belongs to revision rather than to judgment. It shows which sentences read as machine-written to a machine, which is information about the prose and about nothing else. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The public check takes up to 1,200 words, one paper a day, with no account. Human's detector scores the text in memory and de-identifies any personal information for additional training. The sentences worth revising are the ones a student cannot explain out loud, and that test needs no software.

No. An AI-writing detector reads the prose and estimates how much of it looks machine-produced, a judgment about style and structure rather than about sources. The AI reading itself uses no index of the web and no repository of past papers, whatever else the product it ships in may do. Whether a quoted passage carried a citation is outside what that reading can see.

Of the two questions here, Human answers the style one. It reads a piece of writing and reports how much of it reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. It does no plagiarism checking.

Nothing in that design lets a checker of style audit a bibliography. So a paper can come back clean on both instruments and still break the rule. Another can come back marked on both and be entirely a student's own work. The report describes text. The policy describes conduct.