Understanding Your Turnitin AI Detector Report Score

The number in your Turnitin AI detector report is not a plagiarism score and it is not a verdict. It's the percentage of qualifying sentences in your submission that Turnitin's model predicted were generated by AI.

So a 38% doesn't mean 38% of your paper was copied from somewhere. It means the model judged roughly that proportion of the prose it analysed as statistically resembling machine-written text, based on patterns it learned from training data.

That distinction matters enormously, because the two numbers get treated the same way in conversation and they are nothing alike. One points at a source you can go and read. The other points at nothing — only at a probability.

What the Turnitin AI detector actually measures

Turnitin's AI writing indicator breaks a submission into overlapping segments of a few hundred words, scores the sentences inside each one, and averages the results into a single percentage. There is no source document, no match, no link to click.

A few practical consequences follow from how it's built:

How to read the percentage, including the asterisk

The number is easier to interpret once you know which band it falls into.

  1. No score shown. The submission didn't qualify — too short, not enough continuous prose, or an unsupported language. This is not a pass or a clearance.
  2. 0%. No analysed sentence crossed the model's threshold. It means the detector found nothing, not that nothing is there.
  3. 1–19%. Turnitin flags this range with an asterisk because false positives are more common at the low end. Treat these as noise unless something else in the work raises a question.
  4. 20% and above. This is the range where Turnitin reports a document-level false positive rate of under 1% in its own testing. Vendor-run testing on vendor-chosen data is still the weakest kind of evidence available, so read the figure as a claim rather than a measurement.

Even at the high end, "under 1% of documents" is not "under 1% of your students." In a cohort of 400 papers, a sub-1% error rate still means a handful of honest writers can be flagged, and no report tells you which ones they are.

Where the Turnitin AI detector gets things wrong

Every detector has characteristic failure modes. These are the ones worth knowing before a conversation turns into an allegation.

Plain, formulaic prose reads as AI

Detectors key on predictability. Writing that is short-sentenced, cautious and vocabulary-light scores higher, which is why a Stanford study found GPT detectors misclassified non-native English writing far more often than native-speaker writing. Methods sections, lab reports and heavily templated assignments have the same problem for the same reason.

Editing tools shift the signal

Turnitin says it isn't targeting spelling and grammar fixes, but any tool that rewrites whole sentences — a grammar assistant's full-sentence rewrites, a machine translation pass, a paraphrasing service — moves text towards smoother, more predictable phrasing. That is exactly what the model is tuned to notice.

A low score proves nothing either

Paraphrasers and humanizer tools exist specifically to push generated text under detection thresholds, and they often work. This is the asymmetry nobody likes: the report can flag honest writing, and it can wave through text that was generated wholesale.

What to do if your own writing was flagged

Panic-rewriting is the worst response, and running your draft through a humanizer is worse still — it converts a disputed paper into a genuinely altered one.

  1. Ask for the specifics. The percentage, which passages were highlighted, and whether the score carried the low-confidence asterisk.
  2. Produce your process. Document version history, Google Docs revision timeline, dated drafts, handwritten notes, library loans, search history, messages to a tutor. Evidence of how the work was made outweighs a probability score every time.
  3. Offer to talk it through. Explaining your argument, your sources and why you cut a paragraph is something no generated draft can fake convincingly.
  4. Get an independent reading. Running the same text through a different model tells you whether the signal is robust or an artefact of one vendor. You can paste the draft into our free AI text detector and see which sentences it highlights, or use the AI essay detector if the submission is coursework.
  5. Read your institution's policy. Most academic integrity policies require evidence beyond a tool output before a sanction can follow.

If you're the instructor reading the report

The score is the beginning of an enquiry, not the conclusion of one. What's useful in the report is the highlighting, not the headline number: read the flagged passages and ask whether they sound like the student's other work, whether the citations exist, whether the argument holds together.

Our guide to reading an artificial intelligence detector's evidence covers the corroboration checks that actually survive an appeal — invented references, mismatched register, claims the student can't explain.

How the Turnitin AI checker compares to standalone detectors

Turnitin's advantage is integration. It sees the whole submission inside the workflow the institution already uses, with no separate upload.

Its disadvantages are closure and opacity: students generally can't check their own work, the model isn't independently auditable, and you get one number for an entire document.

Standalone tools — GPTZero, Copyleaks, Winston AI, Pangram and others — use different training data and different thresholds, which is why they routinely disagree with each other and with Turnitin on the same paragraph. So do AI detectors work? In aggregate, the better ones separate large volumes of generated text from human text well above chance. On one paragraph by one writer, they are far shakier than their marketing suggests, and agreement between two independent tools is more informative than either score alone.

One more gap: text detectors say nothing about your figures. If a submission includes images that may have been generated, that's a separate check with an AI image detector.

Frequently asked questions

Can students see their own Turnitin AI score?

Usually not. The AI writing indicator is shown to instructors and administrators, and most institutions don't expose it to students. If you want a pre-submission sense of how your prose reads, you'll need a tool you can run yourself.

Is a Turnitin AI detection score proof of cheating?

No, and Turnitin's own guidance says the score should prompt a conversation rather than settle one. It's a statistical prediction about writing style, with no source document behind it. Disciplinary findings need evidence about how the work was produced.

Will using Grammarly make Turnitin flag my paper?

Basic spelling and punctuation corrections are unlikely to matter. Full-sentence rewrite features are a different story, because they replace your phrasing with smoother, more predictable phrasing — the exact pattern detectors score highly. Keep the rewriting light and keep your version history.

Can Turnitin detect ChatGPT text that has been paraphrased?

Sometimes. Turnitin has added paraphrase detection, but heavily reworked generated text can and does slip through, so a clean report isn't evidence of original authorship. If you want to compare readings, our ChatGPT detector highlights passage by passage rather than returning a single figure.

The one thing worth doing now

Turn on version history in whatever you write in, and keep your drafts. A timestamped revision trail is the only thing that reliably answers a detection score — and it takes thirty seconds to set up, long before anyone asks you to prove you wrote your own paper.