Yes. Since April 4, 2023, every eligible submission to Turnitin passes through an AI writing detector that runs inside the Similarity Report and produces a separate, instructor-facing percentage. It doesn't catch everything — short texts, code, and lists are skipped, and low scores are hidden behind an asterisk. Here's the machinery, stage by stage.
This post is the anatomy lesson. If you want the practical view — what happens when you submit and how three kinds of drafts actually fare — read what happens when you submit ChatGPT text. For the evidence on how well any of this works, the full Turnitin review is the deep version.
One report, two detectors
The most useful thing you can learn about Turnitin in thirty seconds is that it runs two unrelated checks and displays them side by side.
The Similarity Report is the original product. It compares your text against a database of web pages, published journals, and previously submitted student papers, looking for matching strings. It answers one question: does this text already exist somewhere?
The AI writing indicator, added in April 2023, answers a completely different question: does this text look like a language model produced it? It doesn't compare your writing against anything in a database. It can't — ChatGPT generates novel text that exists nowhere else. Instead, it examines the statistical texture of the prose itself.
Two checks, two questions, two percentages. Keeping them straight matters more than almost anything else in this topic, and we'll come back to it, because confusing them is how a student ends up celebrating a 3% score that was never about AI at all.
Step 0: what triggers AI processing — and what gets skipped
Before Turnitin's AI model reads a word of your essay, the submission has to qualify. The gatekeeping rules are specific, and they explain most of the "why didn't my document get an AI score?" confusion.
The document needs roughly 300 words of continuous prose. Not 300 words total — 300 words of actual flowing sentences. A submission made of bullet fragments can clear 300 words and still be skipped, because the model was trained on long-form prose and Turnitin doesn't trust it outside that habitat.
The text needs to be extractable. Scanned images of pages, corrupted exports, and unsupported file types never reach the classifier. The Similarity Report has the same constraint, which is why "file processed, no AI score" usually traces back to formatting.
The language matters. The detector was built and validated primarily on English. Submissions in other languages are either not processed or handled with reduced coverage, which is one reason accuracy claims shouldn't be casually extended across languages.
And whole categories of content are excluded by design: computer code, mathematical equations, tables, bibliographies, and bulleted or numbered lists. The model judges prose rhythm, and code has none in the relevant sense. In a flagged lab report, you'll typically see highlighting on the discussion section and nothing on the methods table — that's not the model being selective, it's the gatekeeper excluding non-prose before scoring begins.
If your submission fails these gates, the instructor sees no AI percentage at all — a dash or an error state, not a zero. A missing score and a 0% score are different findings.
Step 1: segmentation — your essay gets sliced
Once a document qualifies, Turnitin doesn't score it as one blob. The text is divided into segments — runs of a few hundred words that overlap with their neighbors, so every sentence gets evaluated several times in slightly different contexts.
Why bother? Because the realistic case isn't a fully AI or fully human document; it's a mixture. A student writes two sections, generates a third, edits everything lightly. Document-level scoring would smear that into a meaningless average. Segment-level scoring is what lets the instructor's view show which passages look machine-generated, as colored highlighting over the text rather than a single number floating in space.
Overlapping windows also smooth out edge effects. A sentence sitting at the boundary between your writing and pasted AI text gets judged as part of both neighborhoods, and the model's calls on it get reconciled across those overlapping looks.
Step 2: classification — what the model is actually measuring
Each segment goes to a classifier trained on large collections of known-human and known-AI writing. It outputs a prediction: the probability that this segment was machine-generated. Segments above Turnitin's internal threshold get marked as AI; the rest pass as human. You never see the raw probabilities or the threshold — that's a vendor decision, tuned (Turnitin says) to keep false positives low at the cost of missing some AI text.
What is the classifier actually looking at? Two families of signal dominate, and they're worth understanding because every mainstream detector uses them. We cover the general theory in how AI detectors work; here's the short version.
Perplexity is next-word predictability. Language models write by choosing likely words, so their output is text a language model finds unsurprising. Human writing is messier — we pick odd words, make lopsided constructions, interrupt ourselves. Run text through a scoring model and machine prose registers as eerily smooth, human prose as spiky.
Burstiness is variation in sentence shape. People write a five-word sentence, then a forty-word one with two subordinate clauses, then a fragment. Models tend to settle into an even, medium-length cadence, sentence after sentence, like a metronome. Low variation reads as machine-typical.
Notice what neither signal measures: authorship. The classifier detects how machine-typical text is, not who typed it. A human who writes highly predictable, evenly cadenced prose — a non-native English speaker trained on formulaic structures, say, or anyone taught the rigid five-paragraph essay — produces exactly the statistical profile the model flags. That gap between "machine-typical" and "machine-written" is where every false positive lives. A 2023 study in Patterns (Liang et al.) made the point empirically: seven GPT detectors falsely flagged an average of 61.22% of 91 human-written TOEFL essays, while judging essays by native-speaking US 8th graders almost perfectly. Turnitin wasn't among the seven tools tested, but its indicator rests on the same statistical foundation the study was probing.
Step 3: the percentage — what the number actually means
The segment verdicts roll up into the headline figure: the share of qualifying prose that the model predicts was AI-generated. A 40% score means the flagged segments amount to about 40% of the text the detector examined — after the gatekeeper discarded your block quotes, references, and lists.
Read that definition again, because two popular misreadings die on it.
A 40% AI score is not "40% confident this was AI." Confidence lives inside the per-segment classification and is never shown. The percentage is a coverage measure — how much of the document got flagged — not a certainty measure.
And it's not "40% chance you cheated." The model has no idea whether your course allows AI, whether the flagged passage was a properly disclosed AI-assisted summary, or that the writer might be a human who happens to write with low burstiness. It reports pattern-matching on statistical texture. Everything about intent, permission, and misconduct is a human's job downstream.
Step 3½: the asterisk — the score Turnitin refuses to show
Here's a design decision that deserves more attention than it gets. When the computed score lands between 1% and 19%, Turnitin doesn't display the number. The instructor sees an asterisk (*) instead.
Turnitin's stated reason: in that low range, its own testing found a higher incidence of false positives, so showing "12%" would lend false precision to a figure the company can't stand behind. The headline claims — 98% accuracy, under 1% false positives — are explicitly scoped to documents where more than 20% of text is flagged. Below that line, Turnitin effectively marks its own output "unreliable."
It's easy to be cynical about detection vendors, and this site usually gives you reasons to be. But the asterisk is genuinely responsible engineering: a company building uncertainty into its interface instead of hiding it in a footnote. The wrinkle is what happens on the receiving end. An asterisk still looks like a flag, and an instructor who doesn't know the policy can read "* " as "AI detected" rather than "score too unreliable to display." The interface concedes uncertainty; whether the human reading it does is another matter.
Similarity score vs. AI score: the confusion that won't die
The two numbers sit in the same report, and they get conflated constantly — in Reddit threads, in syllabi, occasionally in misconduct hearings. Here's the clean separation.
| Similarity score | AI writing score | |
|---|---|---|
| Question asked | Does this text match existing sources? | Does this prose look machine-generated? |
| Method | String matching against a database | Statistical classification (perplexity, burstiness) |
| Evidence type | Shows you the matching source | Shows highlighted segments, no source (none exists) |
| A high score means | Text overlaps published/submitted work | Text is statistically machine-typical |
| A zero means | No database matches found | No qualifying segments flagged |
| Visible to students | Usually yes | Usually no — instructor-facing |
| Low-range handling | Number shown | 1–19% hidden behind an asterisk |
The scores move independently, and the extreme combinations are both possible and common. A fully ChatGPT-written essay typically scores near 0% similarity — the text is novel — while scoring very high on the AI indicator. A meticulous, fully human literature review stuffed with quotations can score 35% similarity and 0% AI. One number can't corroborate the other, because they're not measuring the same thing. "Turnitin score" without a qualifier is an ambiguous phrase, and anyone deciding a misconduct case on the wrong one is deciding it on noise.
What the indicator can't tell anyone
The step-by-step machinery above is real and reasonably sophisticated. It's equally real that the output has hard limits, most of them acknowledged by Turnitin itself.
It can't identify the author. Machine-typical is not machine-written; that's the false-positive gap, and it falls hardest on non-native English writers, students trained in rigid essay structures, technical writers, heavy self-editors, and neurodivergent writers.
It can't name the model. The indicator won't tell an instructor "this was ChatGPT" versus Claude or Gemini — it flags the statistical family resemblance shared by large language models generally.
It struggles most on blends. Mixed human/AI documents are the acknowledged hard case: the Washington Post's April 2023 testing showed Turnitin both missing AI text and flagging human text in hybrid drafts, and Turnitin agreed that blended documents are where the tool is weakest. Turnitin has since said it extended detection toward paraphrased and humanizer-processed text (announced around 2025) [VERIFY current scope], which tells you where the arms race has moved — and also that the vendor considers plain statistical detection insufficient for edited text.
And it can't hold up a verdict alone. Turnitin's own guidance positions the score as the start of a conversation. Vanderbilt University, running the false-positive arithmetic across tens of thousands of yearly submissions, judged even a sub-1% error rate too costly and disabled the indicator entirely in August 2023. How well the machinery performs against every published test is its own story — we've collected every number in Turnitin's accuracy record.
What to do with all this
If you're a student: know that the pipeline exists, that it mostly sees long-form prose, and that its output is a statistical estimate your instructor is told to treat as a starting point. Write in your own voice, keep version history, and read your course's AI policy before touching any tool. If you want to see the same signals a detector sees, you can run a draft through our AI detector, which shows sentence-level readouts of machine-typical patterns — it doesn't promise to predict Turnitin's number, because no external tool honestly can.
If you're an instructor: use the highlighting, not just the percentage. Ask what got skipped, whether the score cleared 20% or hid behind an asterisk, and whether the flagged student sits in a documented false-positive risk group. The machinery is a decent tripwire and a terrible judge. Turnitin says roughly the same thing, in politer language.
FAQ
Does Turnitin detect AI automatically on every submission? On accounts with the feature enabled, yes — the AI check runs alongside every eligible Similarity Report with no separate action by the instructor. Institutions can disable it, and some (including Vanderbilt in August 2023) have.
Why does my document have no AI score at all? It probably failed a gate: under ~300 words of continuous prose, a non-supported language, an unextractable file, or content that's mostly lists, code, or equations. No score means "not evaluated," which is not the same finding as 0%.
What does the asterisk (*) mean on a Turnitin AI report? It means the computed score fell between 1% and 19%, a range Turnitin's testing found too false-positive-prone to display as a number. Turnitin's accuracy claims only cover documents flagged above 20%, so an asterisk should be treated as "unreliable signal," not "AI detected."
Is a 30% AI score worse than a 30% similarity score? They're not comparable — different systems measuring different things. Similarity means text matched existing sources and shows you which ones; the AI score is a statistical prediction with no source behind it. Each needs its own interpretation, and neither is proof of misconduct by itself.
Can Turnitin tell which AI tool was used? No. The classifier detects statistical patterns common to large language models generally; it doesn't attribute text to ChatGPT, Claude, Gemini, or any specific model.
Does Turnitin detect AI-assisted grammar fixes from Grammarly or editors? Ordinary grammar and spelling corrections generally don't create machine-typical statistical patterns across whole segments. Full-sentence rewriting features are riskier, since they replace your phrasing with model-generated phrasing — the more rewriting, the more the text's texture shifts toward what the classifier flags.
Does the AI score affect the similarity score? No. The two checks run separately on the same submission and neither feeds the other. A document can be high on one, both, or neither.
Key facts
- Turnitin's AI writing indicator launched April 4, 2023, as a panel inside the existing Similarity Report (Turnitin).
- Processing requires roughly 300 words of continuous prose; the model is built primarily for English and skips code, lists, tables, and equations (Turnitin).
- Scores of 1–19% are displayed as an asterisk, not a number, because Turnitin found low-range scores unreliable (Turnitin).
- Turnitin's 98% accuracy / <1% false positive claims apply only to documents where more than 20% of text is flagged (Turnitin AI writing FAQ).
- Detectors classify text on perplexity (next-word predictability) and burstiness (sentence-length variation) — they measure machine-typical texture, not authorship.
- Liang et al., Patterns, 2023: seven GPT detectors falsely flagged an average of 61.22% of 91 human-written TOEFL essays; Turnitin was not among the seven but shares the statistical approach.
- The Washington Post (April 2023) found Turnitin weakest on mixed human/AI drafts — a limitation Turnitin acknowledged.
Sources
- Turnitin — AI writing detection FAQ / transparency page (processing requirements, asterisk policy, accuracy scoping).
- Turnitin — product documentation on AI writing report display and instructor guidance.
- Liang, W., et al. — "GPT detectors are biased against non-native English writers." Patterns (Cell Press), 2023.
- Fowler, G. — Washington Post test of Turnitin's AI detector, April 2023.
- Vanderbilt University — statement on disabling Turnitin's AI detection, August 2023.
- Turnitin — announcement on extending detection to paraphrased/AI-humanized text, ~2025 [VERIFY current scope].