Turnitin AI detection, explained

Turnitin's AI writing indicator is a classifier that scores a submission sentence by sentence and reports the percentage it judges likely AI-written. It is not a watermark reader and cannot see any provider's mark. Since August 27, 2025 the report also includes bypasser detection, a second signal trained on the output of humanizer tools, and text it flags counts toward the AI total. Turnitin's own guide says the score is an indicator meant to start a conversation, that short submissions are unreliable, and that bypasser detection is trained and tested for English only. Some universities have switched the indicator off over false positive concerns.

How the score is produced

The model breaks a submission into segments of a few hundred words, scores each sentence on the probability that a language model wrote it, and aggregates. The headline percentage is the share of the text above its threshold. Turnitin publishes the broad method in its detection model guide, including a minimum text length below which no score is given, and notes that the model is retrained as new writing models appear. The August 4, 2026 release merged the two highlight colours in the report into a single set of blue highlights for likely AI-generated text, including text that may have been passed through bypassers or paraphrasers, for English submissions.

What bypasser detection changed

Before August 2025, running model output through a rewriter reliably lowered the score. Turnitin then trained a signal on humanizer output itself, on the theory that rewriting tools have styles of their own. Flagged text is counted as AI-generated rather than reported separately. The feature is available inside the Originality add-on and the AI capabilities add-on for iThenticate 2.0, and only for English; the Spanish and Japanese detectors carry no paraphrase or bypasser detection. Independent testing of the feature has been limited and mixed, and the rewriting tools retrain against it, so results change over time in both directions.

What the score cannot tell you

It cannot tell you which tool was used, whether a person edited the text heavily, or whether a plain-styled or non-native writer simply writes predictably. It cannot be reproduced by the student, because the report is instructor-side. And it cannot be checked against a provider watermark, because Turnitin has no provider keys. Turnitin's guide is candid about the first of these: the score should prompt a conversation, and the instructor should look at the student's process and history.

Why institutions disagree about it

Some universities use the indicator as one input among several and have written policies that forbid acting on the score alone. Others have disabled it, citing false positives and the impossibility of verification. The 2023 Stanford study, which found a 61.3 percent average false positive rate on non-native essays across seven detectors, did not test Turnitin, but it set the terms of the debate, and institutions with many international students have been the most cautious. If you teach, find out what your own policy says before the first flagged paper, not after.

Practical notes

For instructors: check length and language before trusting a score, read the full report rather than the headline number, and ask for process evidence. For students: keep version history on by default, write in a document that records history, and if you are flagged follow the response guide. For everyone: watch Turnitin's release notes, because the model that scored last term's essays is not the one scoring this term's.

Questions people ask

What percentage on Turnitin is considered AI?

Turnitin reports the share of a submission its model judges likely AI-written. It does not define a pass mark and says the score should start a conversation, not decide one.

Can students see the Turnitin AI score?

Usually not. The AI writing report is shown to instructors on licensed accounts.

Does Turnitin detect QuillBot or other paraphrasers?

Turnitin's bypasser detection, launched August 27, 2025, is trained on humanizer output and counts flagged text toward the AI total, for English submissions.

Why did some universities turn off Turnitin AI detection?

Over false positive concerns, particularly for non-native writers, and because a score cannot be verified by the student or the institution.