Turnitin AI detection
at Australian universities, explained honestly.
Turnitin is used across most Australian universities, and its AI writing indicator is now part of how assignments get reviewed. This page explains what that score measures, how academic integrity processes here typically treat it, what TEQSA has said about assessment in an era of generative AI, and where the honest answer is that nobody can give you a number.
What the AI indicator measures, and what it does not
Turnitin's AI writing indicator estimates the proportion of a submission it believes was machine generated. It is separate from the similarity report that Australian students have been familiar with for years. Similarity points at a source you can go and look at. The AI indicator points at nothing external. It is a statistical judgement about writing patterns: how much sentence length varies, how predictable the word choices are, how regular the structure is.
This is why the two scores behave so differently in an integrity process. A similarity match can be examined and either explained or not. An AI score cannot be examined at all, because there is no underlying document to produce. It is a probability, and probabilities can be wrong in both directions on any individual submission.
TEQSA, and why assessment is changing rather than detection improving
The Tertiary Education Quality and Standards Agency has been notably measured on this subject. Rather than recommending institutions rely on detection, its guidance to the sector has emphasised redesigning assessment so that it remains meaningful when generative AI is widely available, including more oral defence, more in-class and supervised work, and more assessment of process rather than only of final artefacts.
That direction is a tacit acknowledgement of something detector vendors are slower to say: detection alone is not a reliable foundation for integrity decisions. If it were, redesigning assessment would be unnecessary.
For a student, the practical implication is that you are increasingly likely to be asked to discuss your own work, and being able to talk fluently about how you arrived at your argument matters more than any score. That is a fair test and it is also the one that protects an honest student best.
The false-positive problem in an Australian context
Australia has one of the highest proportions of international students in the world, which makes the false-positive research directly relevant here rather than a distant statistic. The Stanford study by Liang et al. found detectors flagged 61.3 percent of essays written by non-native English speakers as AI generated, against 5.1 percent for US-born writers.
The mechanism is not bias in the ordinary sense. Detectors measure predictability, and academic writing produced carefully in a second language is frequently more regular and more formulaic than the writing of a confident native speaker. The detector is doing exactly what it was built to do, and the outcome still lands unfairly on a specific group of students.
Our own testing on a different detector found a 1 percent false-positive rate across 100 verified human paragraphs, with one paragraph nonetheless scoring a full 1.00. Averages are reassuring until you are the outlier.
Reading what you are told
| Situation | What it means | What helps |
|---|---|---|
| High AI indicator on an assignment | A model flagged the writing pattern, no source is involved | Draft history, notes, and the ability to discuss your argument |
| Asked to attend an integrity interview | Investigation, not a determination | Bring evidence of process and ask what else is relied on beyond the score |
| Viva or oral defence requested | Increasingly standard, and aligned with TEQSA's direction | Prepare to explain your reasoning and sources in your own words |
| Told the score alone proves misconduct | Inconsistent with Turnitin's own published guidance | Request the written policy your institution applies to detector evidence |
Practical protection for Australian students
Keep a visible drafting trail
Write in something with version history. A document that grew over weeks, with dead ends and revisions, is the most persuasive evidence available to you, and it is decisive in a way no argument about detector accuracy can be.
Be ready to talk about your own work
With TEQSA pushing assessment toward oral and process-based methods, being able to explain your argument, your sources and your reasoning in conversation is becoming the real test. It is also the one an honest student passes easily.
Read your university's academic integrity policy
Every Australian university publishes one and they differ on how detector output may be used. Knowing what yours says before you need it is worth the twenty minutes.
Check your writing before you submit it
If you want to know how your work scores, find out yourself rather than in a meeting. If your own genuine writing scores high, that is precisely when having your drafting evidence organised matters.
An AI indicator score is a probability about writing patterns, not a finding about authorship. TEQSA's direction of travel is toward assessment that does not depend on detection at all, which tells you how much weight the sector itself places on these scores. Keep a drafting trail and be ready to talk about your work, and you are protected in the way that actually holds up.
We publish our own measured detector numbers, including the runs that failed, in the GPTZero pass-rate study. Those figures are measured against GPTZero rather than Turnitin, for the reason explained above.
Frequently asked.
Honestly answered.
Most use Turnitin, though not every institution has the AI indicator enabled and some have deliberately disabled it over accuracy concerns. Your university's academic integrity policy or library guide will say. Do not assume either way.
Its guidance has focused on redesigning assessment for a world where generative AI is available, including more oral and supervised assessment, rather than on relying on detection tools. That emphasis is itself an acknowledgement that detection is not a dependable basis for integrity decisions.
We cannot claim that and we will not. Turnitin publishes no detector that anyone can measure against, so no honest pass-rate figure for it exists from any vendor. Our published numbers are measured against hosted GPTZero and are labelled as such: 73.7 percent of 114 real runs passed, and 16.7 percent failed outright.
Your concern is supported by the evidence. Non-native English writing is flagged substantially more often because detectors penalise predictable prose rather than detecting authorship. Keep a drafting trail, keep your notes, and be ready to discuss your work. Those three things protect you in a way that arguing about the technology does not.
Gather your evidence before responding: version history, notes, earlier drafts, reading records. Ask what evidence beyond the score is being relied upon, and ask for the written policy. Most Australian institutions have an internal appeals process, and student unions frequently provide free advocacy support for integrity matters.