A Teacher Notices the Voice Change; the Draft History Carries the Conversation

Sometimes, though rarely by running the paper through more software, since the published results are split. What gets asked about is an essay that doesn't sound like the last three from the same student, with rhythms and word choices nobody in the class has used before. In a 2023 post, Turnitin told instructors to treat a highlighted sentence as the start of a conversation rather than a conclusion ².

Humanizers are built against the software. The person marking the paper is another matter.

Pangram Labs, a detection company, defined the category in a 2025 post: a humanizer is a tool that rewrites AI-generated text to evade AI detectors ¹. That's the whole product, and the techniques follow from it. One, documented in the same 2025 post, is inserting nonsensical phrases, on the theory that a detector will read gibberish as unlikely to have come from a machine ¹. A marker reads the gibberish too.

Then comes the slow part, and there's nothing technical in it. Someone asks what the argument on page two means, or asks for the file the paper was drafted in, and a rewriting tool supplies neither answer by itself. The artifacts a rewrite leaves in a paragraph are catalogued in what humanized text looks like.

The measurements point both ways, and the loudest numbers come from companies selling something. A 2026 University of Notre Dame study, scored with Pangram 3.2 and GPTZero, ran AI-labeled text through a commercial humanizer and found fewer than 4 percent of the rewrites still flagged, a false negative rate above 96 percent ³. To a student, that makes the score look beatable. To a teacher, it makes a low score worthless as proof of innocence.

Look at the cohort before the headline. Those findings rest on 642 abstracts of 25 to 500 words ³, short academic passages well under the length of a term paper. Nature reports the opposite failure from Pangram's own technical study: when consumer AI was used to substantially modify human-written student essays, the model still called the results fully human 41 percent of the time ⁴. So a detector can miss a rewrite. It can also miss the machine editing in a paper that never went near one.

The plainest sentence on the subject sits on a humanizer vendor's own ethics page. Undetectable AI states that a single AI text detection score alone is not enough to rule out cheating or deception, and that the company has never condoned cheating ⁵. A detection company and an evasion company agree on that much.

The opposite error has its own number. As of September 2026, Human's detector reports an in-house measurement of 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on, with a ceiling on that of 0.498%. The documents were Human's own held-out set, and no independent audit has repeated the count. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.

Whether the tools deliver what the sales page promises, measured, is the question behind do AI humanizers work.

The draft history shows when text appeared and how it grew, which is the question a voice change raises. Google Docs keeps earlier versions of a document, and a writer can name one so it isn't folded into a neighboring version, up to 40 named versions per file ⁶. A humanizer rewrites the words; it does not write the drafting history, and a history that was staged is what the conversation about page two is there to test.

In 2023 a UC Davis senior, William Quarterman, was referred for academic dishonesty after a professor ran his exam through a detector. His answer was the edit history of the Google document, which showed the paper being built over two or three hours, and a demonstration that the professor's tool flagged famous works as AI-generated. He was eventually cleared ⁷.

The record has three holes, and they're worth knowing before anyone leans on it. Version history in Google Docs is visible only to people who can edit the file, so a paper shared as a view-only link hasn't shared its history ⁶. Word keeps version history only for files saved in OneDrive or SharePoint. A document written on a local drive has none to show ⁸. And the owner of a Google Doc can delete unnamed version history, which Google states is permanent and cannot be undone ⁶.

So the record is evidence with conditions, and the conditions are set weeks before anyone asks. Frontiers in Artificial Intelligence draws the same boundary. Its 2026 framework reads process documentation as evidence of engagement, not proof of independent authorship, because process artifacts are not immune to fabrication ⁹.

What a history proves, and what it only suggests, is worked through in Google Docs version history as proof.

It gets flagged. A 2026 Notre Dame study found published abstracts edited by a language model flagged far more often than machine-written rewrites run through a humanizer ³. The authors read that as a sanction risk running the wrong way: under a rule policed mainly by a detector, the permitted edit gets flagged and the concealed draft does not. Carrying that to a classroom is an assumption, since the study measured abstracts and detector flags, not students, sanctions or disclosure.

That's an argument about how courses are assessed. It isn't advice to conceal anything.

A rewritten paper is still a paper whose sentences the student can't account for, and an oral question asks for exactly that account. The Frontiers framework pairs process documentation with an oral defense for that reason ⁹. A student who can explain why the second source was dropped has evidence a rewriting tool does not produce for them.

The defensible position is a real process, recorded while it happens.

What a disclosure line says, and where in a submission it belongs, is set out in an AI disclosure statement.

Start from the draft. A paper written in the student's own sentences from sources the student actually read needs nothing standing between it and the marker. A check afterwards is proofreading. The reading that comes back describes writing that's already there.

The loop that helps is small. Check a draft, rewrite the sentences that read as machine-written, check the draft again. A writing tool in private preview 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, and it never tells a person that their own writing is not theirs. Each pass is about the prose, and the number moves because the prose did. Rewriting a sentence to sound like its author is a different job from rewriting it to beat a classifier.

Length changes what a check is worth. Human's detector declines to answer under 50 words, and between 50 and 149 words it has less to read, so a reading on one short paragraph is weak whatever it says. That's a limit of the instrument. It's also why the unit to check is the whole draft. The free check reads 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.

Keeping the draft history is cheap insurance. A named version in Google Docs survives the merge that folds unnamed ones together ⁶, and a file kept in OneDrive or SharePoint keeps a history a local copy never had ⁸. Neither step takes a minute, and both give the conversation something a rewrite does not.