QuillBot's Paraphraser Rewrites a Draft That Already Exists, and the Course Policy, Then the Campus Code, Decides

QuillBot's Paraphraser hands a paragraph back wearing different words, and it needs that paragraph first. Its own Chrome Web Store listing, published by QuillBot (Course Hero), LLC, describes the Paraphraser as a way to rewrite text to sound clearer, more concise or more natural ¹. The same extension also ships an AI Text Generator that drafts full responses in seconds ¹, and generating a draft is a different act from rewriting one.

That distinction carries everything.

A chatbot asked for 4 paragraphs on the Dust Bowl returns 4 paragraphs nobody wrote. Run those through a paraphraser and they come back reordered and resynonymized, still nobody's. The field has no idea where its input came from, which is why the question is about the paragraph that went in.

One student drafts a body paragraph and runs it through a paraphraser because it reads clumsily. That is editing their own work with a machine. Another pastes a chatbot's answer into the same field. That hides a source.

The keystrokes match. The 2 acts do not.

QuillBot's listing promises a rewrite that adjusts tone without losing the original meaning ¹, so the argument sits where the first draft left it. What a paraphraser moves and what it leaves alone is worked through in what a paraphraser changes and what it leaves alone.

Stanford's teaching guide sorts course AI policies into 4 options, and a paraphrase lands somewhere different under each. They run from yes always allowed through yes but and no but to no never allowed ². The middle 2 carry a condition. Stanford's worksheet lists conditions an instructor can attach, including that students openly explain how and why they used AI tools and that they cite the tools and prompts they use ².

How a paraphrase might land under each, on this article's reading rather than Stanford's:

Silence is the case most students meet.

A course with no AI clause states no permission either way, and the campus policy is the next document to read. The City University of New York names artificial intelligence systems inside its own definition of cheating, wording added in the 2024 revision ³. The page's own explanation says the revision was needed because the 2022 policy, amended and replaced on June 27, 2022, did not address the advent of artificial intelligence ³.

CUNY then goes at the paraphrase question directly. Its policy treats AI-generated content as plagiarism in whole or in part, even when paraphrased, where the AI is not cited as the source ³.

The University of Southern California's academic integrity office writes the same rule from the other end. Material created by a generative AI tool and presented as the student's own work counts as plagiarism whether it was copied verbatim, in near-verbatim form, or paraphrased ⁴.

CUNY names no product. USC names ChatGPT as an example and applies the rule to any generative AI tool ⁴.

Reading the 4 shapes a syllabus AI policy takes in week 1 beats guessing in week 9. Whether undisclosed AI text is plagiarism in the older sense of the word is a separate argument, worked through in what plagiarism means when a chatbot wrote the sentences.

A detector reads the prose in front of it, not the history of who typed it. Pangram Labs, whose detector many schools use, published a humanizer table in August 2025 listing Quillbot at 100.0 percent accuracy on its own humanized test set, and states that Pangram performs above 90 percent on all the notable humanizers it tested ⁵. That figure is the vendor's own measurement of the vendor's own model.

A company measuring itself is still the company.

Nature looked at the wider claim in August 2026. Only the accuracy rates the firms announce from internal testing can be truly current, the reporting found, and those rates are not externally verified ⁶.

The 100.0 percent is a claim with a date on it rather than a property of the tool.

The same Nature reporting carries a number about human-written student essays that consumer AI had substantially modified, which is a heavier edit than a tidy. On that cohort, Pangram's model still labeled the result fully human 41 percent of the time ⁶.

Published evidence on paraphrased text points both ways. Pangram reports 100.0 percent accuracy on Quillbot output in its own test set ⁵. A 2026 University of Notre Dame study found that after a different commercial humanizer rewrote AI-labeled abstracts, fewer than 4 percent stayed flagged by Pangram 3.2 or GPTZero ⁷. Which tool did the rewriting and which detector read it both change the answer.

A percentage on a rewritten paragraph describes prose. It says nothing about who typed it, which is why a flag opens a conversation instead of closing one.

QuillBot's own listing sells two more tools on the same page as the Paraphraser: a Humanizer that rephrases text so it sounds more natural and human-generated rather than created by an LLM, and an AI Detector that checks for the classic tells of AI-generated text ¹. That page is read closely in what QuillBot's own material claims about detection.

Because detectors flag light edits far more often than humanized text, on the one study that measured both. A 2026 University of Notre Dame study of 642 abstracts of 25 to 500 words put the result in one sentence: "Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion." ⁷

That is a fault in the design, not a recommendation.

In that study a light machine edit of an abstract was flagged between 38 and 80 percent of the time, depending on the detector and the year of the abstract ⁷. A humanized draft that no detector flags leaves nothing on the record to review at all.

None of which is fixable before Thursday.

Available before Thursday is a declaration specific enough to be dull: which paragraphs, which tool, what it was asked to do. The study did not test declarations. A line saying a paraphraser was run over the 2nd and 4th paragraphs of a draft the student wrote is much harder to misread than a blanket sentence about AI assistance.

The cost runs the other way as well. A paper rewritten sentence by sentence can read less like its author, and a teacher who has marked earlier work may notice the voice change.

Read the rewritten paragraphs against the originals, side by side, and keep whichever version says what the student actually meant. That is the first of two things worth doing before a paraphrased draft goes in, and neither of them is buying a score. The second is to preserve the drafting record that existed before any tool ran.

Version history does the second job for nothing. A document that arrives in a single paste shows no trail of its own, whatever the truth is, and where that record lives and how it fails is set out in the version history a student can actually show.

A score is not the record.

A reading on the draft is still useful as a prompt to look again at a paragraph. Human reads the prose it is given and reports how much of it reads as machine-written or machine-edited, one paper a day of up to 1,200 words. It measured 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%. That is Human's own in-house measurement as of September 2026, on held-out sets it kept back from training and checked by no outside auditor. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.

The expensive version of the evening is the 11pm search, after the draft is already paraphrased, for something that will reverse the rewrite. That search ends at a humanizer: a different product, making louder promises, read by a marker who already knows how the student writes.