Direct answer
Turnitin uses its own proprietary in-house AI writing detection model, not a licensed third-party detector such as GPTZero or Originality.ai. The model is trained to separate AI-generated academic prose from human writing and reports the share of qualifying text it classifies as AI-generated [1]. Turnitin has stated the detector is built and maintained internally, evaluates text in segments, and only surfaces a percentage once its confidence clears an internal threshold — below roughly 20% confidence it displays *% rather than an exact figure [1].
What AI Detection Model Does Turnitin Use to Flag AI Writing?
Turnitin's detector is a proprietary large-language-model-based classifier rather than a repackaged public tool, and that distinction matters when you are trying to interpret a score [2]. Because the model is trained on patterns statistically associated with machine-generated text, it evaluates writing style, token predictability, and sentence-level coherence rather than searching for copied sources [2]. That is why Turnitin repeatedly frames the AI indicator as separate from its similarity (plagiarism) check — the two reports answer different questions and should never be read as one combined verdict [2].
The model does not tell you which AI tool produced a passage. It will not label text as "ChatGPT" or "Claude"; it only estimates whether a segment looks machine-generated, which is a deliberately narrower claim [2]. Turnitin also publishes accuracy guidance warning that any AI detector, including its own, can produce false positives, especially on highly formulaic or heavily edited human writing [2]. For that reason, the vendor advises institutions not to treat the percentage as the sole basis for an academic misconduct decision [2].
Practically, this means the number you see is a probabilistic signal, not a confession. A high score is strong evidence that the model recognized machine-like patterns; a low score is not proof of authorship, and a *% reading means the model's confidence fell below its own reporting threshold [1][2]. Reading the score with that framing keeps you from over-trusting or dismissing it.
How Does Turnitin's AI Detection Model Decide Which Parts of a Paper Are AI-Generated?
The model works at segment level, not document level. Instead of returning one number for the whole essay, it analyzes successive passages and highlights the specific sentences or spans it classifies as AI-generated, then expresses those highlights as a percentage of qualifying text [3]. That is why a paper that is half human and half machine-written often returns a mid-range figure rather than a clean 0% or 100% [3].
Because scoring is segmented, the location of the highlights carries as much information as the total. A single flagged paragraph inside otherwise unflagged prose looks very different from an entire discussion section lighting up, and instructors are trained to read the distribution, not just the headline percentage [3]. Turnitin's own guidance describes the report as a review aid that should be interpreted alongside drafts, version history, and a conversation with the student — never as a standalone judgment [3].
The segment approach also explains some counterintuitive results. Formulaic human writing — templated lab reports, heavily cited literature reviews, or text that has been run through a grammar tool — can produce machine-like patterns and attract flags, while lightly paraphrased AI text can slip under the threshold [3]. Turnitin acknowledges these limitations directly and positions the model as an indicator that requires human interpretation [2][3].
Can You See the Same AI Detection Model's Output on Your Own Draft Before You Submit?
Usually not through Turnitin itself. Turnitin's standard student workflow does not let you generate your own AI or similarity report unless your instructor has deliberately set up a draft-check assignment that permits it [4]. In many institutions that option is simply not enabled, which is exactly why students look for a way to preview the same model's output before the graded submission window closes [4].
Where a preview is available, the value comes from seeing the identical artifact your instructor will see: the same AI percentage, the same highlighted segments, and the same similarity summary [4]. That lets you revise the specific flagged spans instead of guessing, and it removes the surprise factor on results day [4]. It also gives you a defensible record of what you changed and when, which is useful if a flag is ever questioned [3][4].
The key constraint is fidelity. A generic "AI checker" that runs a different model will not reproduce Turnitin's segment boundaries or threshold behavior, so its number can diverge sharply from the official report [4]. If your goal is to know what Turnitin will actually say, the preview has to come from the same detection model and the same report format [1][4].
If you want to see what Turnitin's own detection model reports on your draft — the same AI percentage, highlighted segments, and similarity summary your instructor receives — turnitin0 runs that preview before you submit, so you can revise the flagged spans with time to spare.
FAQ
Is Turnitin's AI detector the same as GPTZero?
No. Turnitin uses its own proprietary model built and maintained internally, not a licensed third-party detector [1][2]. That is why its percentages and segment highlights do not match the output of unrelated AI checkers.
Does Turnitin's AI detection model say which AI tool wrote the text?
No. The model estimates whether a passage looks machine-generated; it does not attribute text to ChatGPT, Claude, Gemini, or any specific tool [2]. The output is a percentage of qualifying text plus highlighted segments, nothing more.
Why does Turnitin show *% instead of a real number?
Turnitin displays *% when its AI detection confidence falls below its internal reporting threshold, around 20% [1]. Those are low-confidence signals, not a precise score, and should be read as "the model is not confident" rather than "the text is definitely human."
Can I run Turnitin's AI detection on my own draft before submitting?
Only if your instructor enables a draft-check assignment; otherwise the standard student workflow does not provide a self-service report [4]. Students who need a pre-submission preview typically use an independent service that reproduces the same report format.
If my paper is flagged, does that mean I cheated?
Not automatically. Turnitin itself warns that AI detectors can produce false positives and advises institutions not to base misconduct decisions on the score alone [2][3]. The report is meant to be reviewed alongside drafts, version history, and context.