A Zero on a Detector and Humanizer Page Is One Detector's Reading

It rewrites the draft, then scores the rewrite with the checker sitting beside the button. A green number there is one detector's reading of that page's own output, and no more. Pangram Labs, a detector company, defines the category on its own blog: a humanizer is a tool that rewrites AI-generated text to evade AI detectors ¹.

Evade is the vendor's own word for it. What the rewrite changes is a question the page does not answer. One technique Pangram's account names is inserting nonsensical phrases, on the theory that a detector reads gibberish as unlikely to have come from a machine ¹. A sentence shaped to satisfy a classifier wasn't shaped to be read, and a grader notices the difference on the page where it lands: what a humanizer does to writing.

The category is a business. Turnitin told NBC News in January 2026 that it keeps a list of 150 tools that adjust text so a detector won't flag it, some charging as much as 50 dollars for a subscription ⁷, and one academic integrity company counted 43 humanizer sites drawing 33.9 million visits in a single month ⁷.

Because two numbers from two models are two readings, not one fact about the paragraph. Detectors can differ in training data, in how they split a document and in the threshold they answer against; Pangram, for one, scores in segments ⁴. Grammarly's own guide warns that a score its detector returns is no clear indication that a professor's preferred tool will show the same one, though it expects the two to point the same way ³.

The disagreement has been measured. A 2023 study in Patterns ran seven detectors over 91 TOEFL essays written by non-native English speakers, and at least one detector flagged 97.8 percent of them, while all seven flagged 19.8 percent of them as AI ². Averaged over those seven tools, the false positive rate came to 61.3 percent ². Same 91 essays, seven different lists of which ones were flagged.

Where a passage sits can matter too. Pangram's chief executive told Nature in 2026 that its detector scores a document in segments, so a paragraph judged human on its own can switch to AI once it's combined with other text ⁴. That is one way an excerpt pasted into Pangram can score differently from the document it came out of.

Pangram's own technical report, written with the University of Maryland, says Pangram 4 labels humanized text AI-generated 97.67 percent of the time on the commercial humanizers it evaluated, a figure the vendor produced itself ⁵. Nature's assessment of the whole category is that the only current accuracy rates are the ones companies announce from internal testing, and none of those has been externally verified ⁴.

The evidence runs the other way as well. A 2026 University of Notre Dame study rewrote 642 published abstracts with an AI model and ran the rewrites through a commercial humanizer. Of the rewrites two detectors had labelled AI, fewer than 4 percent were still flagged afterwards ⁶. No outside test has reconciled the two. A zero on the humanizer's page is not the reading the paper will get from the tool it meets: used a humanizer and still got flagged.

Three numbers get mixed up in this loop, and only the school's is about the paper that will be handed in. The checker on the humanizer's page scores a rewrite its own vendor produced. The school's tool scores the submitted document. A published accuracy rate describes a test set that researchers built for a study, and that set is neither of the other two.

Human reports, as its own in-house finding measured September 2026 and not independently audited, that whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding.

What that trade looks like from a student's side is free humanizer versus paid tier.

A free checker is still worth running on a draft. It is the wrong tool for the last reading before the paper goes in.

Precision, and sometimes the argument itself. Inserted nonsense is the visible end of what a rewrite can do to a paragraph ¹. Any rewrite made to move a score can also change a key term or damage a citation, and a detector score on either side of the loop reports neither. Those are the places a grader reads for.

Marks get lost before any detector runs. A broken citation, a stray fragment where a reference belonged, a term used slightly wrong: each fails on its own in front of a reader who never opens a checker, and the loop measures none of it.

One study points the wrong way for a student who follows the rules. On AI rewrites of published abstracts, not student papers, light, guideline-compliant AI editing carried a higher modelled risk of sanction than concealing machine drafting behind a humanizer ⁶. Applying that to coursework is an assumption. Read it as an argument about how policies get written, not as a reason to hide anything, and as a case for keeping the draft history that shows how a paper was built.

A rewrite done by hand costs an evening and leaves the paper in the student's own words: how to humanize AI text by hand.

That order is the honest version of the loop, and it costs an evening instead of a click. Run the draft through a checker. Read the sentences it marks, then rewrite those from notes with the flagged version closed, so the new sentence comes from what the writer knows and not from what the old sentence said. Check again afterwards. The second reading is another estimate, not a clearance.

A writing tool is in private preview: it reads a draft with the same detector, colours 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, and it never tells a person that their own writing is not theirs. The thing that runs today is the loop above: check the draft, revise the flagged sentences, check again.

Human is tuned for college-level academic writing and reads sentence by sentence, reporting how much of the document reads as machine-written or machine-edited, with a document verdict of Human, Mixed or AI. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. Up to 1,000 words can be pasted in three times a day, with no account and nothing stored.

Length decides what a reading is worth. Below 50 words Human declines to answer at all. Between 50 and 149 words it has less to read and the reading is weak, so a paragraph of fewer than 150 words pasted in on its own comes back weak, whatever it says. Run as much of the draft as the 1,000-word limit allows, not a lone paragraph.

Which tool an instructor will run is a separate question, and it may not be the one on the humanizer's page: Turnitin's AI checker for students.