Flagged as AI on an Essay the Student Wrote: Save the Drafts, Ask for the Report

A detector score describes text. It doesn't describe a person. It reports how much of the prose matched patterns a model associates with machine writing. Turnitin tells instructors to use the information to initiate a conversation, and says it isn't there to draw a conclusion from. That's the vendor's own account of what its number is for ¹.

Turnitin's own published figures decide how much weight one score can carry.

Length comes first. Turnitin raised its minimum from 150 words to 300 in May 2023, saying accuracy improves on longer text ². A short reflection or a discussion post can fall under that floor, and then no reading is produced at all.

The asterisk comes second. Since 8 July 2024 a document reading under 20 percent gets an asterisk in place of the number, with no score and no highlights attributed ⁹; the 2023 report printed the figure with an asterisk beside it, because Turnitin called the reading less reliable there ². A 14 percent score on an older report and a 74 percent score aren't two points on one scale.

Readings over 20 percent aren't rare, either. Turnitin reported 38.5 million submissions processed for AI writing by 14 May 2023. Of those, 9.6 percent came back above 20 percent AI writing, and 3.5 percent landed between 80 and 100 percent, figures Turnitin says include assignments where AI use was authorized ².

Highlighted sentences carry their own error rate. Turnitin puts it at around 4 percent, or 1 in every 25 highlighted sentences turning out to be the writer's own ¹. Take an essay with 40 highlights. By Turnitin's own figure about 2 of them are wrong, and the figure says nothing about whether the rest are right.

None of it identifies an author.

The University of Pittsburgh's Teaching Center has drawn the line already: its stated position is that detectors are not reliable enough to act as proof, and it has switched the AI detector inside Turnitin off ⁶.

A student curious which sentences of the essay a detector would mark can paste up to 1,200 words at human.olive.is, one paper a day, with no account, and see a verdict of Human, Mixed or AI. 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, taken in September 2026 on its own held-out set and not an independent audit, 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%. A private reading like that still isn't a rebuttal, for the reason set out below.

The page on what a Turnitin AI percentage means goes through what a percentage counts, threshold by threshold.

The version history of the file the essay grew in is evidence that existed before the flag did. It's also the part that can quietly disappear. Open the document the drafting actually happened in, rather than the PDF that was submitted, and save its history the same day the accusation arrives.

In 2023 a professor at the University of California, Davis ran a take-home history midterm through an AI detection tool and referred its author, a senior named William Quarterman, for academic dishonesty ³. He was eventually cleared, and the edit history of the Google document he'd written in was part of his answer. It showed the paper being built across 2 or 3 hours ³.

That trail is easy to lose.

Google's own help page says version history is visible only to people who can edit the file, so a document shared as view-only shows its viewer no history ⁵. The same page says deleting version history is permanent and cannot be undone ⁵. Microsoft says version history in Microsoft 365 works only for files stored in OneDrive or SharePoint ⁸. A draft kept anywhere else needs its record from somewhere else.

None of these takes 10 minutes:

• In Google Docs, click Last edit at the top right to open the version history, then name the current version, so it isn't merged into a neighboring one ⁵.
• Download a copy of the document and email it to a personal address, so a timestamp exists outside the school's systems.
• Screenshot the history panel with its dates visible. Keep the copies somewhere the school can't revoke access to.

A blank history proves nothing either way.

Plenty of students draft on paper or in an offline app and type a clean copy into the submission box at the end. That's an ordinary way to work, and it confesses nothing. It does mean the record has to come from somewhere else: notes, an outline, library loans, the sources themselves, the search history that built the argument.

How a history reads, and what a paste of a student's own notes looks like inside one, is covered in using Google Docs version history as proof.

Ask four things in writing before defending the essay at all: which tool produced the score, what the number was, how long the submission was, and what the course policy says happens next. A request in writing creates a record with a date on it. It also hands the instructor something specific to answer, rather than a denial to weigh.

The length question has a hard edge: a submission under 300 words gets no Turnitin reading at all ².

The report matters more than the percentage.

Turnitin's own guidance is to use a highlighted sentence to initiate a conversation, and it says the highlight isn't there to draw a conclusion from ¹. That sentence is quotable in a polite message. It carries more weight than any student's opinion of the tool, because it comes from the company that sold it.

Asking which tool ran settles something else students often get wrong. One detector's score says little about what another will return. Grammarly sells a detector of its own, and it tells its users that a score from its AI detection is no clear indication that an instructor's preferred detector will show the same thing. In the same breath it says it believes the scores should be directionally aligned ⁷. Neither statement makes a second score evidence about the first.

So a clean result pasted in from a free checker isn't a rebuttal.

Two readings that disagree are two estimates. The evidence that moves a case is the record of how the essay was written. No detector produced that record, and no second detector can overturn it.

The written reply says the work is the student's own, names each tool that was actually used and what it did, attaches the draft evidence and asks for a meeting to walk through the paper. It's one message, sent once, short enough to be read in full, and its argument is about the essay rather than the percentage.

A reply that opens by attacking the detector spends its credibility on the wrong target.

The contents, in the order a reader takes them in:

• One sentence saying the essay is the student's own work.
• The dates the drafting happened, from the version history, with the file shared in a way that carries the history with it ⁵.
• Every tool that touched the draft, named: a spell-checker, a citation manager, a translator, a rewrite button. What each one did, in a clause.
• The notes and sources behind the argument.
• A request to meet and talk through the paper, plus a question about what the course policy says happens next.

Concede nothing that didn't happen. A faster meeting is no reason to.

At the University of North Georgia, a Turnitin flag and the disciplinary hearing that followed put Marley Stevens on academic probation for a year ⁴. She says she used only Grammarly's standard spell-checking and grammar features ⁴. Moira Olmsted, at Central Methodist University, had her grade restored and was still warned that a second flag would be treated as plagiarism ⁴.

Wording that keeps a first message short and specific is set out in what to say to a professor after an AI accusation.

If the flag stands, the question moves from the essay to the process: who decides and on what evidence, and what the student gets to see along the way. Read the institution's academic integrity procedure: it names who gives notice, on what evidence, and where an appeal goes. A detector report is one document in that file, and the decision at the end of it belongs to people.

Ask for the report itself, in full, with every highlighted sentence.

Keep the published false-positive record ready. Bloomberg Businessweek tested two detectors on 500 college application essays sent to Texas A&M University in mid-2022, shortly before ChatGPT was released. Bloomberg says that timing made it effectively certain none of them were machine-written ⁴. The services falsely flagged 1 to 2 percent of them, in some cases claiming near 100 percent certainty ⁴.

1 or 2 essays in every 100, from a pile written before ChatGPT existed.

That figure settles nothing about any one essay. It isn't an argument that detectors never work, either. What it does is explain why a panel is asked to weigh evidence from outside the score, the standard the Pittsburgh Teaching Center already applies when it says a reading is not reliable enough to act as proof ⁶.

The steps an appeal takes, and what a written statement in the file does afterwards, are set out in appealing an AI detection accusation.