Grammarly's Own Page Says Its Humanizer Is Not for Bypassing Detectors

Not where AI use is restricted. Grammarly's humanizer page asks whether the tool will help bypass AI detectors and answers in one line: it is not intended to bypass AI detectors ¹. The same page states that in academic settings, or any context where AI use is restricted or must be disclosed, using an AI humanizer to disguise the use of AI may be considered unethical and could constitute cheating or plagiarism ¹.

That is the vendor's wording, not a critic's.

The page, read in September 2026, also encourages transparent and responsible AI use, and asks for AI assistance to be clearly acknowledged in academic or professional work ¹. The warning itself sits under the free text box rather than beside it.

A campus licence is not permission for the act. The City University of New York's 2024 academic integrity policy defines cheating as the unauthorized use or attempted use of material, information, notes, study aids, devices, artificial intelligence systems or communication during an academic exercise ⁶. The operative word is unauthorized. At CUNY, authorization comes from the syllabus or the course instructor ⁶. An institutional site licence covers the software; the assignment's rule, not the licence, says which features it allows.

The wider split between correcting a sentence and drafting one, under a single product name, is worked through in which Grammarly acts a course policy actually charges.

100.0 percent. Pangram Labs published an updated humanizer benchmark in August 2025 listing twenty tools with a detection rate beside each one, and the Grammarly row reads 100.0 percent ². Quillbot reads 100.0 percent on the same table, DIPPER 97.6 percent, and Undetectable AI sits lowest at 90.3 percent ².

Read the sentence attached to those figures. Pangram's own summary of the table is that it performs above 90 percent on all the notable humanizers it tested ². A vendor measured its own product on its own test set in one month of one year, which is evidence about that detector rather than about detection in general.

That date does real work. A rewriting tool ships changes on its own schedule, and a school buys whichever detector it buys. Quoting a 2025 row as a permanent property of the tool overstates it in both directions: the row describes one company's model against the Grammarly output of that year.

What the newer model publishes is narrower and more specific. Pangram's model card for Pangram 4 describes a separate humanizer score, with a segment flagged as humanized once that score meets a threshold set at 0.91 at release ⁴. A segment can come back carrying a humanized flag beside its AI label, which starts a different conversation for the student whose paragraph it is.

Figures published by the company and by outside testers are collected in what Pangram reports about humanized text.

The humanizer is a paste-in rewriting tool, not the red and blue underlines. Grammarly's own support pages separate them. Traditional, non-generative corrections should typically not move Grammarly's own AI detection score, while writing agents and the generative assistant used to meaningfully rewrite sentences and full paragraphs will raise it ³. Two acts, one brand.

So the answer splits on what the tool was handed. A student who pastes a ChatGPT draft into the humanizer is asking one machine to rewrite another machine's sentences. What comes back is nobody's writing. The authorship problem comes first, and no detector threshold changes it.

Grammarly states that it currently offers no dedicated feature for reducing the AI score in a document ³. A company that offers a humanizer and also warns against using it to disguise AI use is drawing a line between rewriting for readability and rewriting for concealment. The line is thin in practice and the page still draws it.

One more sentence from the same guide is worth keeping. Grammarly tells students not to view any score from its AI detection as a clear indication that a professor will see the same score in a different tool ³. Two detectors can disagree, and the reading that counts is the instructor's.

Because evasion sometimes works, and the rule holds either way. A 2026 University of Notre Dame study ran AI-labelled text through Undetectable AI and found that fewer than 4 percent of the rewrites stayed flagged, a false negative rate above 96 percent ⁵. That was measured on AI rewrites of 642 abstracts of 25 to 500 words, scored by two detectors, so it describes short academic passages rather than whole papers ⁵.

One commercial tool, one length band, one study. The same paper reports something harder to file away: honest AI editing carried a higher sanction risk than humanizer-assisted evasion ⁵. In that study, an abstract lightly polished by a model was flagged far more often than a machine-written one run through the humanizer ⁵. Carrying that to a whole paper is an assumption. That is an argument for a fairer process. Nothing in it recommends the humanizer.

CUNY treats unattributed AI-generated content as plagiarism, including content that has been paraphrased, and asks that the AI be cited as the source ⁶. A humanizer pass is paraphrasing with the attribution stripped out. No detection rate touches that.

What a school actually sees after a commercial rewrite, and what the outside testing covered, is set out in what schools see after an Undetectable AI pass.

Start from the notes rather than from the sentence on the screen. A paragraph that came out of a model and then out of a rewriting tool stands two removes from the student's own thinking, and swapping further words does not close the distance. Rewriting from an outline and from a memory of the argument produces sentences a writer can defend in a five-minute conversation. That conversation is what a flag leads to.

A draft can be read before it is handed in. Human takes up to 1,000 words at human.olive.is and marks, sentence by sentence, which parts read as machine-written or machine-edited, three times a day, with nothing stored and no account. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. The marks work as a map of where to start rereading.

Check, revise, check again.

Sentences worth rebuilding are the ones carrying a model's shape: even length, a hedge at the front, three parallel clauses in the middle. Each one goes back to the note that prompted it.

Revising to move a number is the humanizer's job performed slowly by hand. Revising because a sentence says something other than what the writer meant is different work, and it is the version that holds up when somebody asks about paragraph four.

The habits that separate the two, and what a hand rewrite cannot fix, are in how to rewrite AI text by hand without a tool.