Accused After Grammarly: Bring the Draft History, Not a Second Score

Nothing gets edited. Whatever record exists gets preserved first. A Google Doc keeps its own version history, but only people with edit access can browse it ⁵, so a paper shared as view-only has shared none of it. Naming a version stops nearby edits being merged into one another ⁵. Deleting the unnamed history is permanent, Google states on the same page ⁵.

Editing the file that afternoon stacks a new version on top of the ones that matter.

Word behaves differently, and the difference catches people out at the worst moment. Microsoft keeps version history only for files stored in OneDrive or SharePoint ⁶, so an essay saved to a laptop's own drive has no version history to produce. Other records may still exist: earlier attachments, printouts, a phone's notes. A draft typed into a phone and pasted in at the end leaves no revision record either.

The next minute goes to establishing which Grammarly features were switched on.

Grammarly says the generative features are available on demand rather than on by default, and that a user can switch generative AI off in account settings ². A settings screenshot shows whether they were switched off before the draft, not whether they were used. Taking one on the day the email arrives still beats reconstructing the same thing three weeks later in front of a panel.

Once the record is safely stored, what a history establishes and what it does not is worked through in reading a Google Docs version history as evidence.

Grammarly answers that on its own support site. Red and blue underline corrections should not normally move its AI Detector score, while its writing agents and its generative assistant do raise one when they meaningfully rewrite sentences and full paragraphs ¹. A student who accepted comma fixes and a student who asked for a rewrite both say they used Grammarly. They did different things.

One of those acts is editing and the other is drafting, and only the second writes sentences.

A magazine test in 2024 put a number on the gap. Reporters at Bloomberg Businessweek ran a human-written essay through Grammarly's rewriting features, and work that had passed as 100 percent human-written came back reading as 100 percent machine-written, while ordinary spelling and grammar fixes barely moved the reading ⁴. The same investigation put two other detectors over 500 college application essays submitted to Texas A&M University in mid-2022, before ChatGPT was released. It found 1 to 2 percent falsely flagged, some of them with near-total certainty ⁴.

The same reporting followed Marley Stevens, a student at the University of North Georgia. After a Turnitin flag, a disciplinary hearing found she had cheated and she was put on academic probation for a year, by her account. She said she wrote the assignment herself using only Grammarly's standard spell-checking and grammar features ⁴.

Her account is the shape of the accusation on this page.

Two smaller facts from Grammarly's own pages save time. Grammarly tells students not to read a score from its own AI detection as an indication of what an instructor's preferred detector will report ¹, so a reassuring number there is no clear guide to the number in the gradebook. Grammarly also states that it offers no dedicated feature for reducing an AI score in a document ¹. Worth knowing, before a search for one turns a defensible position into a second accusation.

That same split, read from the other side of the desk, is what a teacher is weighing when Grammarly comes up.

A version history settles that a draft was built over time, and it does not settle who thought of the sentences. A 2026 framework in Frontiers in Artificial Intelligence states that process documentation should be read as evidence of engagement, not proof of independent authorship, because process records are not immune to fabrication ⁹. A file that grew over hours is still often the best record a student has.

It has worked in public at least once. Tech & Learning reported the case of William Quarterman, falsely accused of using AI. He produced the edit history of the Google document he wrote his exam in, and the timestamps showed the paper being built over two or three hours ⁷.

Grammarly keeps a second record, and it is narrower than the name suggests. Grammarly Authorship distinguishes text typed directly into a document from text pasted from a browser-based source, and from text pasted from a source it cannot see, such as a private browsing window ³. Where tracking was on, it also separates text generated by Grammarly's AI, text rephrased with its on-demand AI, and text edited with its non-generative corrections ³. In Google Docs and Microsoft Word it does not track writing activity by default, so the record exists only where tracking was switched on before the draft was written, or where the draft was written in Grammarly's own editor ³. It describes where text came from and how it was edited, not what anyone meant by it. A student who drafted in one app and pasted the result into another lands in the pasted column having done nothing wrong.

The phrasings that keep the meeting from turning into an interrogation are set out in what to say to a professor after a flag.

By asking what the number is a rate of. Turnitin puts its sentence-level false positive rate at around 4 percent, and says that means a 4 percent likelihood that any one highlighted sentence is human-written ⁸. A highlight marks a place to look rather than a sentence shown to be machine-written.

The document-level figure on the same page is under 1 percent, and it holds only for papers the model already reads as 20 percent or more AI writing ⁸.

Turnitin's chief product officer said in June 2023 that false positives are more common in documents the tool reports as under 20 percent AI writing ¹¹. Grammarly's statement is about its own detector's score, and the same page says that score may differ from Turnitin's ¹. Neither company publishes where a corrections-only paper lands on Turnitin's scale, so the low band is a reason to ask which rate applies, not a place the paper is known to sit.

A low percentage, then, is not a smaller accusation.

The most useful sentence to bring comes from Grammarly, the maker of the tool at the center of the accusation: no detector available today can reliably and conclusively determine whether AI was used ². That is not a student's opinion about detection. It is the published position of a company that sells one of these tools, on a support page anyone in the room can open and print.

Running the paper through three more detectors produces three more estimates, and no source here measures how often detectors err on the same paper. Two readings that agree are still two estimates; two that disagree leave open the question they were meant to close. A low number from a free checker carries no standing in a meeting. A high one produces panic about a tool nobody in the room is even using, and panic is what sells a humanizer to a student who did nothing wrong.

At least one detector publishes where its own thresholds sit. Human works on college-level academic writing and reports how much of a document reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. Anything under 50 words goes unscored, and a reading on 50 to 149 words is treated as weak, which covers a good deal of short coursework.

Three settings mark what a percentage is being compared against: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. Pasting in up to 1,200 words, one paper a day, with no account, returns a sentence-level reading and a verdict, and nothing beyond that. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Human's detector scores the text in memory and de-identifies any personal information for additional training.

Human's own in-house measurement, as of September 2026, records 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 figure comes from Human's own held-out sets rather than from an independent audit. A rate across documents is still not a verdict on one paper, which is the shape of the whole argument on this page.

What Turnitin's own numbers say about a lightly edited draft is taken up in the Turnitin side of a Grammarly case.

The syllabus, rather than any campus-wide rule, settles what a student was allowed to use, and one university writing center tells students to consult each one because course policies differ ¹⁰. A clause permitting spelling and grammar support answers the accusation on its own terms. A clause banning all machine assistance turns the underlines into a policy question instead of a detection one.

Those two questions get confused constantly, and separating them does most of the work.

Whether the writing reads as machine-edited is a measurement, and measurements carry error rates. Whether a rule was broken is a reading of a document an instructor wrote, and that document answers it.

One more item belongs in the folder: the assignment prompt itself, plus any email in which the instructor described which tools were allowed.

A student who followed a written instruction has a different case from a student who guessed, and instructors do not always reread their own prompts before a meeting.

If the decision goes the wrong way, the route afterwards is procedural rather than technical: how an AI detection accusation gets appealed.