Artificial Intelligence
When Turnitin Flags Your Essay as AI, the Score Opens a Process and Evidence Closes It
The email arrives with a number in it. Picture a student whose essay, written over three weekends, has been marked as containing a high share of AI-generated text, and the lecturer would like to talk. It is easy to read that number as a verdict, and easy to respond with panic and a temptation to start rewriting honest work.
The published positions point somewhere calmer. A detector score is a prompt for questions, not proof of misconduct. The institution has to hold evidence beyond the score, and the student’s strongest response is calm, in writing, and built on a working trail of drafts, notes, sources and version history, together with the ability to explain the argument in their own words. That is our position. The rest of this article sets out what the score is, what a vendor claim and an independent study can and cannot tell a student, and four steps to take this week. It is general information, not legal or academic advice, and every institution sets its own process.
What the score is, and what it is not
Turnitin’s guide for instructors, a page that carries no date, describes the AI writing indicator in specific terms. It examines prose in long-form writing, with a minimum of 300 words and a maximum of 30,000. The percentage is the share of the submission the model judges likely to be AI-generated or AI-paraphrased, and the model also looks for text altered by a paraphrasing or bypasser tool such as QuillBot. It does not reliably detect non-prose such as bullet points, tables or annotated bibliographies. Scores from 1 to 19 per cent appear as an asterisk rather than a figure, and the guide itself reports “a higher incidence of false positives when the percentage is between 0 and 19”.
Who sees the report matters as well. The University of Melbourne’s academic integrity page on Turnitin and AI writing detection, also undated, says the indicator and report are accessible to staff and not visible to students, although an instructor can share a PDF copy of the report. The same page explains that the percentage “represents the amount of text that it calculates as 98% or more likely to have been generated or paraphrased by AI” and calls the result “a prompt for further investigation”. That is one university’s account of its own practice, and other universities differ.
The question students ask first is the one the percentage cannot answer: which number is safe? No source we found sets a safe percentage, and Turnitin’s guide does not offer one. Each institution’s policy decides what a given score triggers, and some decide nothing on a score alone. What cannot be known from the outside is where your own university draws its line.
The check a reader can run: open the unit’s assessment guide and the institution’s academic integrity policy, and find the sentence that says how AI detection is used.
What the vendor says, and what the independent evidence found
Two claims circulate, and they are often set against each other. They deserve a ruling.
Turnitin’s product page for its originality checking states a false positive rate of less than 1 per cent for documents containing more than 20 per cent AI-generated content. We report that as the vendor’s claim, not as a verified finding, and it carries a stated condition: it describes documents above that level of AI content. The vendor’s guide for instructors is open about the band below it, where false positives are more frequent, which is why the lowest scores are hidden behind an asterisk.
The independent evidence addresses a different problem. Liang, Yuksekgonul, Mao, Wu and Zou posted a study to arXiv in April 2023 (third version July 2023, also published in Patterns) finding that several widely used GPT detectors “consistently misclassify non-native English writing samples as AI-generated” while accurately identifying native samples. The authors caution against using such detectors in evaluative or educational settings.
| What the vendor says | What the independent study found |
|---|---|
| A false positive rate of less than 1% for documents with more than 20% AI-generated content (Turnitin’s product page for its originality checking, a vendor claim, not independently verified) | Several widely used GPT detectors “consistently misclassify non-native English writing samples as AI-generated” (Liang et al., arXiv 2304.02819, first posted 6 April 2023, revised 10 July 2023) |
| The detection “may not always be accurate” and “should not be used as the sole basis for adverse actions against a student” (Turnitin Guides, Using the AI Writing Report, undated) | The authors caution against use in evaluative or educational settings (same abstract) |
| “A higher incidence of false positives when the percentage is between 0 and 19” (same Turnitin guide) | Native English samples were accurately identified, so the errors fell on one group of writers (same abstract) |
Our reading is that the two do not contradict each other. They involve different detectors, different test conditions and different questions, and a figure for one condition says little about a particular student’s essay. We have not tested any detector and state no error rate of our own. Together, though, they support one rule: a score is not evidence of who wrote something.
The vendor deserves its fair point here, because its guide says the same thing. The sentence is “sole basis”, not “no basis”, and that difference matters. A common view among students online is that Turnitin has admitted it is wrong and therefore the institution has no grounds. The guide does not say that. It says the output takes “further scrutiny and human judgment”, and a high score the writer cannot explain still invites more questions.
Step 1: ask for the report and the policy
Ask in writing for the report and for the policy that applies. At the University of Melbourne an instructor can share the PDF, so the request is reasonable there, and at other institutions it costs nothing to ask. Then ask two more things: what evidence the institution relies on beyond the score, and what it would accept from you in reply.
This step answers the fairest complaint against the whole process. A score can feel like a verdict to a student who is never shown the report, and some students report being asked for proof of authorship without being told what the concern rests on. Asking is the remedy, and a polite written request creates a record that the question was raised.
Practice differs between institutions. Curtin University’s update, published on 4 September 2025, said the AI writing detection feature in Turnitin would be turned off from 1 January 2026 across all campuses and study periods. It gave its reasons as trust and clarity in academic culture and assessments that are secure, fair, relevant and future-ready, and it confirmed that text-matching checks and integrity procedures continue. The University of Melbourne, for one, keeps the indicator available to staff only. You cannot assume that your own campus matches a friend’s.
The check a reader can run: find the name and section number of the policy that applies, and quote it back in your email.
Step 2: a meeting is not yet an allegation
The University of Melbourne states that being asked to discuss your work “is not an allegation of academic misconduct”. A discussion of this kind typically covers how the argument developed, the sources used, and sometimes a request for drafts or notes. The same page says “an AI writing detection report alone is not sufficient evidence for an allegation” and, on attendance, that “you are not required to participate in or attend a discussion if you choose not to”.
Three cautions belong beside that. It is one university’s published position, so check your own. A formal notice is a different thing from an invitation to talk, and a formal notice should always be answered by the date given. And the right to decline is not the same as advice to decline: our reading is that a short, prepared conversation is often the quickest way to settle a question, though the decision is the student’s and an adviser at the student union can help weigh it.
Students enrolled through TAFE or another registered training provider sit under that provider’s own integrity policy, not a university’s, and the same applies to those in Free TAFE places, which differ from state to state. The wording quoted above will not carry over automatically, so look for the equivalent document at your provider.
A check you can run before any meeting: write down in two sentences what the essay argues and in one sentence why you chose your main source.
Step 3: gather your working trail
The strongest answer to a score is a record of how the work was made. Most students already hold more of it than they realise, scattered across devices and folders.
Collect what exists: early and later drafts, notes and outlines, saved sources with the pages or highlights you used, and the version history in Word, Google Docs or OneNote. Keep local copies of all of it. Do not delete anything while a question is open, including drafts you consider embarrassing, because every draft shows how the work developed.
Version history deserves a fair caution. One caution raised in student discussions is that a steady typing history on its own proves little, because it can be reproduced. We found no institutional source on this, so treat it as a reason to keep drafts and notes as well, not as a rule. Our reading is that ideas that changed, sources that were added and sections that were reworked show more than typing history alone. Pair that with the ability to explain the argument aloud, including why the essay is structured as it is and what changed between drafts.
Two points on AI use itself. Where a task allowed AI assistance, the University of Melbourne’s page says that use had to be acknowledged and cited, so check what your assessment allowed and make sure the acknowledgement was there. And Turnitin’s guide documents that the model looks for text altered by paraphrasing tools. Heavy use of those tools is one documented reason for a flag. If a rewriting or grammar tool touched the work, say so and keep the before-and-after versions.
To see where you stand today, open the document properties and version history for your last assignment now, and see how far back it goes.
Step 4: reply in writing, and ask for support
Keep the reply calm and factual. State that you wrote the work, list what you can supply, and ask what the institution would like to see. If you quote Turnitin’s caveat, quote it exactly: “sole basis”, not “no basis”. If English is an additional language for you, say so, because that is the pattern the independent study identified.
Contact the student union’s advocacy service early. The University of Melbourne’s page points to the advocacy service run by its student union, and other institutions may offer an equivalent through a guild or association. An adviser can help you read the policy and weigh your options while you are under stress.
Four things to avoid, each with a reason:
- Do not delete drafts. They are the evidence, and a gap where they should be raises its own questions.
- Do not paste the essay into a free checker or a “humaniser”. The University of Melbourne warns that some free tools can be very inaccurate, that many are aimed at exploiting student anxieties and claim to find AI in wholly human-written work so they can sell you “humaniser” tools, and that pasting an essay into one gives the work away.
- Do not rewrite honest work to “sound less like AI”. We found no source that supports the idea, and changing the document mid-process muddies the trail.
- Do not ignore a formal notice. Whatever the merits, deadlines in the notice usually run whether or not you reply.
If the outcome goes against you
Review and appeal steps exist and differ by institution. Ask in writing for the policy section that governs review, and for the deadline, because those windows can be short. The student union’s advocacy service can help prepare a submission, and where the institution has an independent review office, it can explain what it can and cannot examine. No outcome is predictable from a score, and none is predicted here.
Keeping your own paper trail
The habit that helps on any assignment is a boring one. Keep drafts and sources in a folder named for the unit. Write in a tool that keeps its own history. Ask what AI use is allowed before you start, not after a flag. A first-year student who has just come through the ATAR scaling process has been used to a system in which the marks are produced and ranked for them. Coursework integrity runs the other way, because the trail is yours to keep.
The verdict stands: the score opens the process, and evidence closes it. A student who holds the drafts and can explain the argument is in a far stronger position than one holding only a protest.
The check a reader can run: open the assessment guide for your next task, find the line on AI use, and copy it into the top of your working file.
Sources: Using the AI Writing Report, Turnitin Guides (https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report); Turnitin and AI writing detection, University of Melbourne Academic Integrity (https://academicintegrity.unimelb.edu.au/plagiarism-and-collusion/advice-for-students-regarding-turnitin-and-ai-writing-detection); Update on Turnitin AI-Detection Tool, Curtin University, 4 September 2025 (https://www.curtin.edu.au/news/oasis-news/update-on-turnitin-ai-detection-tool/); GPT detectors are biased against non-native English writers, Liang et al., arXiv 2304.02819 (https://arxiv.org/abs/2304.02819)