Flagged by Turnitin for AI: What to Do Next

A Turnitin AI flag is a prediction, not a finding of misconduct

The percentage on your screen is a model's estimate of how much of your prose resembles machine-generated writing. It is not a verdict, and Turnitin does not present it as one. The company's instructor guide states that the AI writing detection model "may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student." Its FAQ goes further: "Turnitin does not make a determination of misconduct; rather, it provides data for the educators to make an informed decision based on their academic and institutional policies."

That matters, because it means you are not arguing against the company. You are quoting it. Detection is probabilistic by design, and the vendor documents that in its own help pages.

So the productive question is not how to disprove a number, but what the rest of the record shows about how the document came into existence. That record is usually stronger than the score, and most of it already sits on your laptop. For the accuracy figures themselves, we have covered how accurate Turnitin's detector actually is separately. This post is about procedure.

What the percentage counts, and what it leaves out

Turnitin's FAQ defines the figure precisely: the indicator shows "the amount of qualifying text within the submission that Turnitin's AI writing detection model determines was likely generated by AI or likely generated and modified by an AI paraphraser or bypasser." Qualifying text is narrow. The same page specifies "only prose sentences, meaning that we only analyze blocks of text that are written in standard grammatical sentences and do not include other types of writing such as lists, bullet points (short non-sentence structures), or other non-sentence structures."

Then comes the sentence most students never see, also from Turnitin's FAQ: "This percentage is not necessarily the percentage of the entire submission."

Picture a 2,000-word report in which 600 words sit in a results table, a bulleted method list and a reference list. Turnitin scores the remaining 1,400 words of prose. If the model flags 350 of those words, the indicator reads 25 percent. That is a quarter of the qualifying text, not a quarter of the file you uploaded. When you are asked to account for a number, knowing its denominator is the first correction worth making.

The AI percentage is also independent of the similarity score, and Turnitin's guide notes that AI writing highlights are not visible in the Similarity Report at all. Two systems, one screen.

Why you might be looking at an asterisk instead of a number

If your report shows an asterisk rather than a figure, that is deliberate. Turnitin's guide states that for scores above 0 percent and below the 20 percent threshold, "no score or highlights are attributed," and the result "is now indicated with an asterisk (*%)." The stated reason is plain: "To avoid potential incidence of false positives." The same guide notes that reports generated before July 8, 2024 may still display a numerical score under 20 percent.

The University of Bristol's Digital Education Office announced the change to staff in July 2024, noting Turnitin had also raised the maximum submission length for AI detection to 30,000 words.

Turnitin's own August 2023 whitepaper explains where the threshold came from. A document is labelled AI-written when more than 20 percent of its sentence-level scores clear the model's threshold. The paper states: "Based on tests conducted by us, we've determined that in cases where we detect less than 20% of AI writing in a document, there is a higher incidence of false positives. Hence, the 20% document proportion cutoff as well as the predetermined model threshold were chosen to keep document level FPR below 0.01 (1%)."

An asterisk is therefore not a suppressed bad score. It is the vendor declining to show a number it considers too unreliable to act on. If one is being treated as evidence against you, raise that politely and in writing.

What your instructor can see that you cannot

Turnitin's FAQ is explicit: "The AI writing detection indicator and report are not visible to students," and "only instructors and administrators are able to see the indicator." You may be the only person in the conversation who has not read the document being discussed.

What they see is more detailed than a percentage. The AI Writing Report opens with the overall figure, then an interactive submission breakdown bar mapping flagged passages across the pages of your file. Highlights are selectable, so clicking one jumps to the exact sentences the model scored. That specificity works in your favour: specific passages can be discussed and explained. A bare number cannot.

Ask for the report: the request that changes the conversation

Turnitin's FAQ notes that "with the PDF download feature, instructors can download and share the AI report with students." It is the most actionable line in the documentation, and almost nobody quotes it.

So your first move is a short written request: ask for the PDF of the AI writing report, and ask which specific passages are highlighted. Send it by email rather than raising it after class. Email creates a timestamped record and gets you the specifics before any meeting.

Four lines are enough. "Dear Dr [Name], I understand my submission for [assignment] returned an AI writing indicator in Turnitin. Could you send me the PDF of the AI writing report, and let me know which passages are highlighted? I am happy to meet at your convenience to go through it." Keep it to that. This message is a request for a document, not the place to make your case.

Whether they share it is up to them and to institutional policy, but asking is reasonable, not defensive. How to handle the conversation itself is a longer subject, and we have written a full guide to what to do when your essay is flagged as AI.

Build your authorship file before you reply

Do this before you write back, not after. The strongest answer to a probabilistic flag is a concrete account of how the document was built. Most of that evidence already exists, and some of it expires, so collect it today.

1. Google Docs version history. Open the file and click the "Last edit" link at the top, or use File, then Version history. Google's support documentation notes that you need edit permission on a file to browse its earlier versions, and that you can create named versions so revisions are not merged together. Screenshot the timeline and export a copy.

2. Word version history, with an important catch. Microsoft's support page states it directly: "Version history in Microsoft 365 only works for files stored in OneDrive or SharePoint in Microsoft 365." Signed in with a personal Microsoft account you can retrieve the last 25 versions; with a university account the number depends on how your library is configured. A .docx that only ever lived on your desktop has no version history at all, which is exactly why the next item matters.

3. Everything else carrying a timestamp. Outlines, notes photographed on your phone, library and database search records, drafts emailed to yourself or a writing centre, draft submissions inside your LMS, comments from a friend who read it. Any two that bracket the drafting period are worth more than one perfect screenshot.

4. Your sources, in your head. Be ready to explain why you chose a particular paper, what you disagreed with, what you cut and why. Fluency about your own reading is hard to fake, and in a room it persuades more than any screenshot.

5. A one-page timeline. Dates, what you wrote when, and an honest inventory of every tool you used and what for: spellcheck, reference manager, grammar checker, translation, anything generative. A timeline that omits something and is later contradicted does more damage than the original flag ever could.

One hard line runs through all five: assemble, do not manufacture. Backdating a file or producing drafts after the fact turns a defensible situation into fabrication of evidence, which institutions treat as more serious than the original allegation.

Four things in Turnitin's documentation that explain odd scores

Short submissions behave in an all-or-nothing way. Turnitin's FAQ says that "in shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap," with the consequence that "some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated." The minimum length for any report is 300 words of prose. A discussion post that scrapes over that floor is the least stable input the system accepts.

Grammar checking and generative drafting are treated differently. Turnitin's FAQ states that its detector "is not tuned to target Grammarly-generated spelling, grammar, and punctuation modifications," and that in its own testing such changes "were not flagged as AI-written by our detector" in most cases. The same answer carves out "content generated by Grammarly's generative AI-powered features, including draft generation, paraphrasing, summarizing," which it says "will likely be flagged as AI-generated." If you ran a grammar checker, say so. The distinction is documented and it works in your favour.

Drafting in another language and translating carries a real cost. Weber-Wulff and colleagues, publishing in the International Journal for Educational Integrity in 2023, tested twelve public detectors plus two commercial systems, Turnitin and PlagiarismCheck. Overall accuracy on human-written English was 96 percent; on human-written documents machine-translated into English, the paper reports that "the accuracy dropped by 20%", a phrasing it leaves open between percentage points and a relative fall. The authors' explanation was direct: "machine translation leaves some traces of AI in the output, even if the original was purely human-written." If that describes how you work, it belongs in your timeline.

Paraphrasing tools are an explicit target, not a loophole. Turnitin's FAQ states that its detection covers "likely AI-generated content that may have been modified using a word spinner/AI paraphrasing or bypassing tool to evade detection," and that the company is "unable to disclose the names of these tools." Anyone selling you certainty about how a rewriter performs against Turnitin is selling something the documentation does not support.

How an academic integrity process actually runs

Procedures vary by institution, so read yours rather than a summary. A published example gives the shape. The University of Southern California's Office of Academic Integrity describes an administrative review beginning with notice sent to the student by university email, with the instructor copied, then a meeting with office staff, a determination of responsibility, referral to a review panel where more severe sanctions are possible, and an appeal route.

The standard of proof USC publishes is "preponderance of the evidence (i.e., what is more likely to have occurred)." UK institutions usually phrase the same idea as the balance of probabilities. Neither is the criminal standard.

That cuts both ways. A lower bar means doubt alone will not carry you. It also means the decision turns on the weight of the whole record, and a probabilistic score with nothing corroborating it is thin under any standard, which is exactly what Turnitin tells instructors. Two things are close to universal whatever the local wording: you are entitled to know the specific allegation, and you are entitled to respond to it.

Some universities switched the detector off, and why that matters

Vanderbilt University published a post titled "Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector" on 16 August 2023, with arithmetic any student can follow. Vanderbilt submitted 75,000 papers to Turnitin in 2022, before the AI detector existed. Applying the 1 percent false positive rate Turnitin claimed, the university calculated that had the tool been running, "around 750 student papers could have been incorrectly labeled as having some of it written by AI." The University of Pittsburgh followed weeks later, its University Times reporting that the Teaching Center had disabled Turnitin's AI detection "effective immediately," judging current detection software "not yet reliable enough to be deployed without a substantial risk of false positives."

None of this is a trump card, and produced as one it lands badly. In the peer-reviewed testing behind much of the scepticism, Turnitin scored highest of the fourteen systems examined and recorded no false accusations across the test documents. Institutional doubt is context that belongs alongside your evidence, not in place of it.

What if the work really was AI-assisted?

Policies are not uniform, so read the exact clause rather than assuming. Some institutions permit generative AI for brainstorming but not drafting. Some permit drafting with disclosure and a citation. Some prohibit it outright. Some say nothing useful at all, which is its own kind of problem.

If you used AI in a way your policy allows, say so plainly, show where, and point at the disclosure requirement you followed. If you used it in a way your policy does not allow, get advice from a students' union, an academic adviser or an ombudsperson before you reply to anything. What is never worth it is inventing a paper trail. That turns one allegation into two, and the second is the worse of them.

Whether editing AI output is itself misconduct has no single answer across institutions, and we have set out the honest version of that argument in our piece on whether using an AI humanizer counts as cheating. The short version: the tool is not the ethical fact. What your institution permits, and whether you are straightforward about what you did, are.

Before the next deadline, check your own draft

The asymmetry running through all this is that the tool judging your work is one you cannot run. Turnitin sells to institutions, not students, so unless your school offers a self-check portal, your instructor sees your score before you do.

The practical answer is to check your prose against a detector you can run yourself. Our AI detector is free, needs no account, and handles up to 300 words per check. It is not Turnitin, and that needs saying plainly: no two detectors agree, and a low score with us does not predict a low score there. What it does tell you is whether your writing carries the statistical signature detectors respond to, and which passages carry it. That is the part you can act on.

If a paragraph reads as machine-written and you wrote every word, the usual causes are fixable without touching your argument: uniform sentence length, stock transitions, and vocabulary belonging to nobody in particular. Our AI word checker flags the stock phrasing specifically. Where AI assistance was permitted and your policy allows you to edit and submit the result, genuine rewriting means restructuring the argument and adding the detail only you have, which is what our humanizer is for, rather than swapping synonyms.

The habit worth building today, whatever happens with this flag: write in a document with version history switched on, from the first sentence. It costs nothing, and it is the only evidence that assembles itself while you work.