How does Turnitin detect AI? The indicator explained

Turnitin's FAQ describes four steps, and each one removes something from the text before a number appears.

Turnitin states that the document-level score "is the percentage of sentences that are above the threshold" and that it does "not display scores below 20% in the AIW report" (Turnitin, AI writing detection capabilities FAQs, read 9 November 2026). The percentage is therefore a share of qualifying prose, not of the file. A lab report that is mostly tables could show a high percentage from a few paragraphs, and Turnitin says as much: the percentage "is not necessarily the percentage of the entire submission".

In October 2026 Turnitin updated the model architecture for English submissions to improve accuracy at transition points where text changes between human and likely AI-generated content, read 9 November 2026. Scores on documents already submitted do not change unless the document is resubmitted.

Turnitin's explanation has two layers. The plain-language layer says language models generate "the next highly probable words" while human writing is "inconsistent and idiosyncratic", and that its classifiers are trained to detect those differences in word probability (Turnitin, AI writing detection capabilities FAQs, read 9 November 2026).

The technical layer is more careful. Turnitin states that its current system is a deep learning model based on the transformer architecture, trained on its in-house corpus of academic writing, and that it "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics", read 9 November 2026. Its outputs come from many learned patterns together, which Turnitin says is why "individual predictions may not always be explainable in simple feature-by-feature terms".

That design choice matters for anyone who has read that detectors "measure perplexity". Turnitin says its model does not, at least not as a named feature, so a rewrite aimed at one statistic has no predictable effect on its score. The guide on what AI detectors measure covers the detectors that do publish named signals, and does Turnitin detect humanized text covers the paraphrase question on its own.

Turnitin states one accuracy figure, and it is a false positive figure. Turnitin keeps its false positive rate "under 1% for documents with over 20% of AI writing", tests on "over 700,000 additional academic papers that were written before the release of ChatGPT" before every model release, and accepts that holding that rate means missing some AI text, so that a document shown as 50 percent AI "could contain as much as 65% AI writing" (Turnitin, AI writing detection capabilities FAQs, read 9 November 2026).

Turnitin also names the text most likely to trigger a false positive: "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas", and warns that in documents of only a few hundred words the prediction is "mostly all or nothing" because a single segment is scored, read 9 November 2026.

Weber-Wulff and colleagues included Turnitin among the 14 detectors in their 2023 study and found that no tool reached 80 percent accuracy across 54 documents, with Turnitin among the lowest for false positives on their human texts (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Vanderbilt University switched the indicator off in August 2023, reasoning that a 1 percent false positive rate across its 75,000 yearly submissions would be about 750 papers (Vanderbilt University, guidance on AI detection, 16 August 2023, read 9 November 2026). Both are findings about 2023 models; the guide on whether AI detectors work sets them beside the current vendor claims.

Students do not see it. Turnitin states that the indicator and report "are not visible to students", that only instructors and administrators see them, and that an instructor can download the report as a PDF and share it, read 9 November 2026. The feature is licensed through Turnitin Originality or as an iThenticate add-on, institutions can switch it off, and Turnitin processes a submission for AI only when the institution has the feature enabled. Any website offering a "Turnitin AI score" to the public is not Turnitin.

On what a score should do, Turnitin is explicit. It "does not make a determination of misconduct"; it "provides data for the educators to make an informed decision based on their academic and institutional policies"; and the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors", all read 9 November 2026. No detector score should be the only evidence in a misconduct case.

The Turnitin pages set out a colour code for the indicator: blue with a percentage when processing succeeded, an asterisk when AI was detected below 20 percent, grey dashes when the file could not be processed, and an exclamation mark for a processing error. A reader who sees an asterisk is looking at a document Turnitin itself declined to score.

Because it cannot reproduce them. Turnitin is licensed to institutions, not sold to individuals, so any Turnitin reading would come from one instructor's account on one date and could not be checked by anyone else. The published method runs every humanizer output through five detectors that are open to the public, GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, records each one's own label and the date, and publishes the raw files.

What this site does report about Turnitin is what Turnitin publishes, with the date it was read, and that is the page you have just read. The Turnitin evidence page lists twelve humanizers on the same plan fields with no Turnitin result for any of them, and the guides hub holds the rest of this series.