How Does Turnitin Detect AI? What the Score Means

How does Turnitin detect AI? It splits the long form prose in a submission into overlapping segments of roughly five to ten sentences, scores each sentence in context for how likely it is to be AI written, and reports the share of qualifying sentences that cross a threshold as a percentage. Scores from 1 to 19 percent show only as an asterisk because false positives are more likely there (Turnitin AI writing FAQ; Turnitin white paper).

Turnitin itself calls the result an indicator, not proof. Its own guide says the model "may not always be accurate" and "should not be used as the sole basis for adverse actions against a student" (Using the AI Writing Report). This guide explains what Turnitin checks, how the percentage is built, what the asterisk means, and how students and teachers should respond to a flag. Everything here comes from Turnitin's own documentation, checked on October 24, 2026.

Short answer

Turnitin uses a trained classifier, not a list of telltale words. Its FAQ describes a transformer based deep learning model trained on Turnitin's own large corpus of academic writing plus AI generated text. The model is "not explicitly programmed" to measure single signals such as perplexity or burstiness; it learns many patterns at once, which is why Turnitin says a single prediction cannot always be explained one feature at a time. In October 2026 Turnitin updated the English model to work better at the transition points where a document switches between human and likely AI text (release notes).

The output has three parts: an overall percentage, highlights on the sentences the model flagged, and a breakdown into "AI generated only" and AI generated text that was likely changed by a paraphraser or bypasser tool.

What text Turnitin checks and what it skips

The AI score only covers qualifying text, which Turnitin defines as prose sentences in long form writing such as essays, dissertations and articles. During processing the model drops strings it predicts are not real sentences: lists, headers, references, data tables, and text that looks like formulas or code. Turnitin also says the model does not reliably detect AI text in poetry, scripts or code, or in short form writing such as bullet points, tables and annotated bibliographies.

A submission also has to meet file rules before it gets an AI score at all:

Two practical effects follow. First, the percentage is a share of the qualifying sentences, not of the whole file, so a lab report full of tables can show a high percentage on a small amount of prose. Second, Turnitin warns that a document mixing writing types can show a "disparity between the percentage and the highlights."

How the percentage is calculated

Turnitin's white paper and FAQ describe the same pipeline:

1. Segment. The text is cut into windows of a few hundred words, about five to ten sentences each. The windows slide forward one sentence at a time, so each sentence is seen in several overlapping windows.
2. Score. For each window the model outputs a number between 0 and 1, where 0 means very unlikely to be AI written and 1 means strongly plausible.
3. Average per sentence. A sentence's score is a weighted average of the windows it appears in.
4. Threshold. Sentences above a tuned threshold are labeled AI; the rest are labeled human.
5. Report. The document score is the percentage of evaluated sentences above the threshold.

Turnitin says it tunes these thresholds to keep the false positive rate under 1 percent for documents with more than 20 percent AI writing, and it tests every model update on more than 700,000 academic papers written before ChatGPT was released. The tradeoff is stated openly: to keep false positives low, the model misses some AI text, so "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing."

The overlapping windows also explain a known weakness. In short documents of only a few hundred words there is a single segment and no overlap, so Turnitin says the prediction is mostly "all or nothing," and mixed writing can be flagged as entirely AI.

Why low scores show an asterisk

Turnitin's testing found more false positives when the score falls between 1 and 19 percent. Since July 2024 the report shows no number and no highlights in that range, only an asterisk (*%). A score of 0 means nothing was flagged; 20 to 100 percent shows a number and highlights. Reports generated before July 8, 2024 may still show numbers below 20 (Using the AI Writing Report).

So an asterisk is not a hidden accusation. It means the model found a small amount of text above its threshold and Turnitin judged that number too unreliable to display.

False positives and how to respond

Turnitin lists the kinds of human writing most likely to be misread: text "without a lot of structural variation," text that "literally repeats itself," and text paraphrased without developing new ideas. In May 2023 it also changed its logic after seeing more false positives in generic opening and closing sentences. Independent research on other detectors has found higher false positive rates for some groups of writers, such as nonnative English speakers; our guide to whether AI detectors are accurate covers those studies.

If you are a student whose own writing was flagged:

• Ask which sentences were highlighted and what the percentage was. Turnitin's FAQ says the report is not visible to students by default, though instructors can download it as a PDF, and its newer report guide describes instructors making the report available to students.
• Bring process evidence: drafts, notes, sources and document version history.
• Offer to talk through the argument and how you wrote it. Turnitin says the decision on misconduct rests with the instructor and the institution's policy, not the tool.

If you are a teacher reviewing a flag: treat it as a reason to look closer, check the highlights against what you know of the student's writing, and use a conversation and drafts before any penalty. Our step by step method in how to detect AI writing shows how to stack evidence instead of relying on one score.

Does paraphrasing beat it?

Less often than it used to. Turnitin added detection of AI word spinners in December 2023 and of "bypasser" (humanizer) tools in August 2025, and both now feed the same percentage, for English submissions only. Turnitin will not name the paraphrasers it tests against. Grammar fixes are a different matter: Turnitin says that in its tests, spelling and grammar changes from Grammarly and similar checkers were mostly not flagged, but text from Grammarly's generative features, such as drafting or paraphrasing, will likely be flagged.

No detector is a fixed target. Turnitin notes that scores can change as the model is updated, but an existing report only changes if the work is resubmitted. For how Turnitin compares with consumer detectors, see GPTZero vs Turnitin and Pangram vs Turnitin, and for the broader picture of methods, what is AI detection.

FAQ

Can Turnitin detect ChatGPT?

Turnitin says its English model detects text from a long list of models, including GPT-4o, the GPT-5 family through GPT-5.4, several Claude, Gemini and Llama models, DeepSeek V3.2 and Grok 4.1. Detection is probabilistic, so it can miss AI text and can flag human text.

What AI percentage is too high on Turnitin?

Turnitin sets no cutoff for misconduct. Scores from 1 to 19 percent are hidden behind an asterisk because they are less reliable, and Turnitin says any score should start a conversation, not decide a grade. Your school's policy decides what happens next.

Can students see the Turnitin AI score?

Not by default. Turnitin's FAQ says only instructors and administrators see the indicator, though instructors can download the report as a PDF to share, and Turnitin's report guide describes instructors making it available to students.

Does Turnitin check code or bullet points for AI?

No. The model skips lists, headers, references, tables, formulas and code, and Turnitin says it does not reliably detect AI in non prose such as poetry, scripts or code. Turnitin also says it is not pursuing code detection at this time.

What should I do if Turnitin flags my own writing?

Stay calm and gather evidence: drafts, notes, sources and version history. Ask which passages were highlighted, explain how you wrote them, and point out that Turnitin itself says the score should not be the sole basis for action.