Turnitin AI Detector: How It Works

What the Turnitin AI detector is

Turnitin added an AI writing indicator to its Similarity Report. It shows an overall percentage of a submission that the model predicts was generated or modified by AI writing tools, and links to a report that highlights the sentences behind it. The Similarity score and the AI percentage are independent of each other: one counts text that matches other sources, the other estimates AI writing. Turnitin says it does not make a determination of misconduct; it provides data for educators to weigh (Turnitin FAQs).

The report splits the flagged text into two categories: text likely generated by a large language model, and text likely generated and then revised with an AI paraphrasing tool or word spinner. Since August 2025 the first category also covers text that may have been modified by an AI “bypasser”, the tools sold to make machine text look human (release notes).

How it works, according to Turnitin

Turnitin describes a classifier. It takes the sentences of a submission, groups them into overlapping segments, gives each segment a value between 0 and 1, lets each sentence take its segment’s value, pools the values where segments overlap, and aggregates the sentence values into the document score. It says the model is transformer-based and trained on its own corpus of academic writing, including AI-generated text and writing by groups such as second-language learners, and that it is not explicitly programmed to check individual signals such as perplexity or burstiness.

For English, Turnitin lists the models its detector can recognize, including recent versions of GPT, Gemini, Claude, Llama, Mistral, DeepSeek and Grok and the tools built on them, and says it will keep expanding the list. It updates the model from time to time: its release notes record changes to improve recall while keeping false positives low in October 2025 and February 2026, and its FAQ says that in July 2026 it merged a multi-model ensemble into a single model while keeping the false-positive rate under 1%.

If you have read how AI detectors work in general, this is the same design: overlapping windows, a learned score, and a combination step. Our guide to how AI detectors work explains what follows from it.

What the number and the symbols mean

• A percentage. The share of qualifying text that the model judges likely AI-generated or modified. It is not the share of the whole file: only long-form prose sentences count, so lists, bullet points and code are left out, and bibliographies are excluded.
• An asterisk instead of a number. When the model finds AI writing below 20%, Turnitin shows an asterisk (*%) and no percentage or highlights, to avoid false positives. This has applied since July 2024.
• Gray dashes. The report could not be produced: the submission was made before the feature existed, it doesn’t meet the file requirements, or something went wrong.
• Old reports stay old. When the model changes, reports already made are not rescored. A submission has to be resubmitted to get a score from the newer model.

What a file needs to be checked

These figures are from Turnitin’s file requirements page. Turnitin adds that accuracy increases with more text. In documents of a few hundred words its prediction is mostly “all or nothing”, because there is a single segment and no overlap to average over, so a mix of human and AI writing can be flagged as entirely AI-generated.

Who can see the result

• Instructors and administrators see the indicator and the report. Administrators can switch the feature on or off, and since February 2024 they can view aggregated AI scores for their institution.
• Students cannot see the indicator or the report. An instructor can download the report as a PDF and share it.
• No repository. Turnitin says AI detection does not add a paper to a repository. The results are kept as part of the Similarity Report.

Because the report is not visible to students, there is no official way for a student to run Turnitin’s detector on their own draft, unless an instructor shares a report.

What Turnitin says about false positives, and what others found

Turnitin says it aims to keep the false-positive rate, meaning fully human-written text identified as AI-generated, under 1% for documents with more than 20% AI writing, and that before each model release it tests on over 700,000 academic papers written before ChatGPT. It also says the low rate has a price: some AI writing is missed, so a document it scores at 50% AI could contain as much as 65%.

It names the kinds of text that can cause false positives: content without much structural variation, text that literally repeats itself, and text that has been paraphrased without developing new ideas. In an earlier post it gave a second number, a sentence-level false-positive rate of around 4%, meaning a 4% chance that a highlighted sentence is human-written. It said these errors are more common in documents that mix human and AI writing, and that 54% of them sit right next to actual AI writing (Turnitin blog).

Outside studies are less reassuring. Weber-Wulff and colleagues tested 12 public detection tools and two commercial systems, Turnitin among them, in 2023, and concluded that the tools were neither accurate nor reliable. Liang and colleagues found that seven widely used detectors flagged, on average, 61.22% of essays by non-native English speakers as AI-generated. Turnitin says its training data deliberately includes second-language learners and English users from non-English-speaking countries.

Vanderbilt University disabled the feature in August 2023. Working from Turnitin’s then claim of a 1% false-positive rate, it noted that the 75,000 papers it had submitted in 2022 could have produced about 750 wrongly labeled papers (Vanderbilt). Our guide, are AI detectors accurate?, sets these and other findings side by side.

If your work is flagged

• Remember what the score is. Turnitin itself says it does not determine misconduct and that the percentage should not be the sole basis for action.
• Ask to see the report. Note whether it shows a percentage or an asterisk, which sentences are highlighted, and how long the document is. Wrongly highlighted sentences tend to sit next to real AI writing, and short documents are judged “all or nothing”.
• Gather evidence of your process. Drafts, version history in Google Docs or Word, outlines, notes and the sources you used show how the work developed better than any score can.
• Be ready to explain the work. Turnitin’s own guidance for instructors suggests asking a student to explain what the essay is about, the process they followed and their key ideas. If you wrote it, you can.
• Know the tools you used. Turnitin detects text that was written by AI and then revised with an AI paraphrasing tool (QuillBot is its example), and says so in a separate category. Check your institution’s policy on AI and paraphrasing tools, and disclose what it asks you to.
• Use the process. Ask what the appeal process is and who decides. If it isn’t clear, a student advocate or ombudsperson can help.

If you are reviewing a report

• Use the score to start a conversation, not to reach a conclusion. Turnitin’s own advice is that there is no “right” or “target” score and that highlighted sentences are areas of interest.
• Weigh the length of the document, whether the writer’s English is plain or learned as a second language, and whether the flagged passages are at the start or the end, where Turnitin has said generic introductions and conclusions caused false positives.
• Compare with the student’s drafts and earlier work, and let the student explain their process before deciding anything.

Free ways to check, and what they can’t tell you

Turnitin makes its detector available to institutions and says the report is not visible to students, so there is no free way to run it yourself. Free detectors run different models. They can disagree with each other and with Turnitin’s result, so treat one as a second opinion on how your writing reads, not as a preview of what your instructor will see.

Whichever tool you use, the most useful evidence is your own process: the drafts, notes and sources that show how the work came together.

About this guide

FreeAIDetector.org is not affiliated with Turnitin, and Turnitin is a trademark of its owner. This guide reports what Turnitin says publicly and what independent sources found, as checked on September 30, 2026. Turnitin changes its product and its documentation often, so check its current pages before relying on a detail.

Sources

• Turnitin: AI writing detection capabilities FAQs. How the model works, what the indicator shows, who can see it, and Turnitin’s false-positive claims.
• Turnitin: File requirements for an AI Writing Report. The length, size, file type and language requirements.
• Turnitin: AI writing detection model (release notes). The dated history of changes: categories, the 20% asterisk rule, bypasser detection and model updates.
• Turnitin: Understanding the false positive rate for sentences. The sentence-level false-positive rate and Turnitin’s advice on using the indicator.
• Vanderbilt University: Why We’re Disabling Turnitin’s AI Detector. The university’s reasons and its false-positive arithmetic.
• Weber-Wulff and colleagues (2023): Testing of detection tools for AI-generated text. An independent test of 12 public tools and two commercial systems, Turnitin among them.
• Liang and colleagues (2023): GPT detectors are biased against non-native English writers. How seven detectors treated essays by non-native English speakers.

Facts and figures were checked against these sources on September 30, 2026. If something is out of date or wrong, tell us.