# What AI Detector Does Turnitin Use? Its Own Model and Score, Explained · ScriptGrain

> Turnitin runs its own in-house AI writing detector. Only instructors see it, scores under 20% show as *%, and Turnitin says the score must not be the sole basis for action against a student.

Canonical: https://scriptgrain.com/blog/what-ai-detector-does-turnitin-use

# What AI Detector Does Turnitin Use? Its Own Model and Score, Explained

*By Jack Stovell [X](https://x.com/StovBuilds) · [Instagram](https://www.instagram.com/stovbuilds) · [LinkedIn](https://www.linkedin.com/in/jackstovell/) · 2026-10-10 · Guides*

Short answer: Turnitin wrote its detector itself and runs it inside its own platform, next to the similarity check it is known for. The result appears as the AI Writing Report. Only instructors and administrators see it by default, and Turnitin's guidance is that the score is one data point, never a verdict on its own.

## The short answer: Turnitin's own detector

No licensed third-party model sits underneath, and it is not a rebadged version of anything you can paste text into on the open web. Turnitin trains and updates the model itself, with separate models for English, Spanish and Japanese. Its file requirements list English, Spanish, Japanese and Arabic as supported languages.

Turnitin's FAQ says the English model can detect content from GPT-4o, GPT-5, o1-mini, Gemini, Claude, LLaMA, Mistral, DeepSeek, Nova and Grok models, plus tools built on them. That list changes as new models appear, so check the FAQ for the current version.

The report also has entry conditions. A file needs at least 300 words of prose in a long-form format, no more than 30,000 words, and must be a .docx, .pdf, .txt or .rtf under 100 MB. Bibliographies are excluded from the analysis.

## What Turnitin's AI report shows

Two things: a percentage, and highlighted text.

The percentage covers what Turnitin calls qualifying text, meaning prose sentences in long-form writing. Turnitin's FAQ is explicit that this "is not necessarily the percentage of the entire submission." Lists, tables and short fragments may not count.

The highlights show which passages drove the figure. Since August 2025, the AI-generated category also includes text the model thinks was produced by an LLM and then reworked with an AI paraphraser, word spinner or "bypasser" tool. Turnitin names Quillbot as an example of a word spinner.

Students do not see any of this unless someone shows them. Turnitin's FAQ says the indicator and report "are not visible to students", though an instructor can download the report as a PDF and share it.

The highlights make the output look precise, as if the tool had pointed at exact lines. In practice it is a model's estimate of where AI-like text sits. It shows where the model's suspicion landed. It cannot show how a sentence was written. For the mechanics, see [how AI detectors work](https://scriptgrain.com/reference/how-ai-detectors-work).

## What a score like 20% or 30% means

A score of 20% means the model judges roughly a fifth of the qualifying prose as likely AI-generated. That is different from 20% of your paper being generated, and very different from proof that you cheated on a fifth of it.

Twenty per cent is also the line Turnitin draws itself. Since July 2024, scores from 1% to 19% show only as an asterisk (*%) with no number and no highlights. Turnitin's reason is that results in that range "are less reliable and false positives are more likely." Reports made before that change may still show a number below 20%.

A 30% score clears the threshold, so you see a figure and highlights, and an instructor may look harder. It is still an estimate. Turnitin's review guidance says "no tool can replace the educator's judgment", and its model notes say a score "should not be used as the sole basis for adverse actions against a student." What decides the outcome is context: your drafts, your earlier work and the assignment itself.

## How accurate is it? Vendor claims and independent tests

Start with the vendor. In March 2023, Turnitin reported a document-level false positive rate under 1% for documents scoring over 20% AI, tested on 800,000 papers written before ChatGPT. Its sentence-level false positive rate on the same test was about 4%. Turnitin also accepts that keeping false positives low means "we might miss some AI written text." These figures come from the company selling the product, so quote them, but keep going.

Usage data shows how often the detector fires. By 14 May 2023, Turnitin had processed 38.5 million submissions. Of those, 9.6% scored over 20% AI and 3.5% scored 80% to 100%.

Now the independent evidence. Weber-Wulff and colleagues tested 14 detection tools in 2023, and all of them scored below 80% accuracy. Turnitin ranked first. Accuracy was 96% on plain human text and 74% on unmodified AI text. It fell to 42% after light human edits and 26% after paraphrasing. Turnitin was the best of the 14 and still missed a lot.

Walters, also in 2023, compared 16 detectors. Only three performed strongly on both AI and human text: Copyleaks, Turnitin and Originality.ai. That is a genuine point in Turnitin's favour.

Then there is the institutional response. Vanderbilt University switched off Turnitin's AI detector in August 2023. It submitted roughly 75,000 papers a year, so a 1% false positive rate implied about 750 students wrongly flagged. Vanderbilt also said Turnitin gives no detailed information on how the detector works, which makes the claims hard to check.

One more study matters for fairness. Liang and colleagues at Stanford (Patterns, 2023) found that seven public detectors were biased against non-native English writers. Turnitin was not among the tools tested. Its FAQ says its training data took account of second-language learners, but that is the vendor's claim. If this applies to you, read [AI detectors and non-native English writers](https://scriptgrain.com/blog/ai-detectors-and-non-native-english-writers). Every figure, with its source, is in the [AI detector accuracy reference](https://scriptgrain.com/reference/ai-detector-accuracy).

So Turnitin is among the better detectors, and it still makes mistakes. Its own documentation says so.

## Why public 'Turnitin checkers' are not Turnitin

Search the phrase and you will find plenty of sites offering a "Turnitin checker". None of them is Turnitin. Its detector runs inside Turnitin's own system, it is not offered as a public service, and even students at subscribing institutions cannot see the AI indicator.

Those sites run some other detector. Their number has no reliable link to the number on your instructor's report, so a low score on one of them tells you nothing about Turnitin, and neither does a high one.

If you want a general check on your own writing for generic AI-style phrasing before you share it, ScriptGrain's [AI detect](https://scriptgrain.com/ai-detect) will flag those patterns. It does not predict Turnitin's score, and nothing here suggests it does.

## Can Grammarly or light editing trigger it?

Turnitin's FAQ answers the Grammarly question directly. The detector "is not tuned to target Grammarly-generated spelling, grammar, and punctuation modifications." Text produced by Grammarly's generative AI features, though, "will likely be flagged." So correcting your own writing is a different case from having a tool write or rewrite it. The full picture is in [does Grammarly get flagged by Turnitin](https://scriptgrain.com/blog/does-grammarly-get-flagged-by-turnitin).

The editing finding from Weber-Wulff runs the other way. It is about AI text that a person then edited, which detectors caught far less often (42% against 74%). It says nothing about how grammar software affects human writing.

Whatever the tool, a detector judges the finished text. It cannot see how you worked.

## If your work is flagged: what to do

A flag is an estimate, and you are entitled to respond to it.

1. Read your institution's academic integrity process. Find out who reviews a flag and what counts as evidence.
2. Gather your drafting history: version history, notes, outlines, earlier drafts and sources you used.
3. Ask to see the report, including the highlighted passages. Instructors can download and share it.
4. Point calmly to Turnitin's own guidance that the score should not be the sole basis for adverse action, and to the published accuracy limits above.
5. Be ready to talk through your argument and sources. If you wrote it, you can explain it.

For a fuller walkthrough, see [how to prove you didn't use AI](https://scriptgrain.com/blog/how-to-prove-you-didnt-use-ai).

## Questions

### Which AI does Turnitin detect?

Turnitin's FAQ says its English model can detect content from GPT-4o, GPT-5, o1-mini, Gemini, Claude, LLaMA, Mistral, DeepSeek, Nova and Grok models, and from tools built on them. Since August 2025 it also flags AI text that has been run through an AI bypasser tool.

### Is 20% AI detection bad?

On its own, no. It means the model judges about a fifth of the qualifying prose as likely AI-generated. Turnitin calls the score one data point and says it should not be the sole basis for adverse action. Scores from 1% to 19% show only as *%, because Turnitin says false positives are more likely there.

### Is 30% Turnitin score high?

It is above Turnitin's 20% threshold, so you get a number and highlights, and an instructor may look harder. Turnitin's own claim of under 1% document-level false positives applies to this range. Even so, about 1% at scale is still a lot of students, and a score remains an estimate.

### Does Turnitin use GPTZero or another public detector?

No. Turnitin built its own detector and runs it inside its own platform. GPTZero and the other public checkers are separate products with their own models.

### Can a student check their own work with Turnitin before submitting?

Not directly. Turnitin says only instructors and administrators can see the AI indicator, though an instructor can download the report and share it. Sites selling a "Turnitin checker" do not run Turnitin's model.

## More from the Journal

- [Wikipedia's Signs of AI Writing: What the List Says and Where It Fails](https://scriptgrain.com/blog/wikipedia-signs-of-ai-writing)
- [What AI Detector Do Universities Use? Tools, Policies and Limits](https://scriptgrain.com/blog/what-ai-detector-do-universities-use)
- [Is the JustDone AI Detector Accurate? What the Tests Show](https://scriptgrain.com/blog/is-justdone-ai-detector-accurate)

[More from the ScriptGrain Journal](https://scriptgrain.com/blog)