Turnitin AI detection
Turnitin's AI detection is a separate indicator inside the Similarity Report, estimating how much of a submission was likely written by a large language model. It is not the similarity score, and Turnitin says the two do not influence each other. The number is the share of qualifying text, meaning prose sentences in long-form writing, that the model judged likely AI-generated, or likely AI-generated and then run through a paraphraser. It is not a confidence level, not a share of the whole document, and not a finding of misconduct: Turnitin's own guide says the score "should not be used as the sole basis for adverse actions against a student". Only instructors and administrators see it, and only where the institution licenses the feature and turns it on.
How the AI indicator differs from the similarity score
For most students, Turnitin is the AI detector for essays they meet first, because it arrives inside the plagiarism check a school already runs rather than as a tool anyone chose. The two numbers it reports measure different things.
The similarity score is a matching problem. Turnitin compares the submission against its collection of student papers, web pages and publications, and reports how much matches something already there. Every match has a source document on the other end.
The AI indicator is a prediction with no source to point at. Turnitin describes the process: it splits the submission into sentences, groups them into overlapping segments, and scores each segment from 0 to 1 for how likely it is machine-written. Each sentence inherits its segment's score, and those scores pool into the document figure. Turnitin says the model is not explicitly programmed for signals such as perplexity or burstiness, and that a prediction is not always explainable feature by feature.
A high similarity score with no AI flag means matching text, and a match can be quotation, a reference list or a stock phrase as easily as copying. An AI flag with low similarity is the reverse: predicted machine writing that matches nothing Turnitin holds. The highlights sit apart, so somebody has to open the AI writing report to see which sentences were marked.
What the percentage actually means
The percentage is a proportion of the qualifying text, not of the whole document: prose sentences inside paragraphs of a longer piece. Turnitin's guide to the AI Writing Report says the model does not reliably detect AI writing in non-prose, naming poetry, scripts and code, and short-form writing such as bullet points, tables and annotated bibliographies. So a lab report that is half tables is scored on a much smaller pool than the number implies, and Turnitin says a document mixing writing types shows a gap between the percentage and the highlights.
The rest of the mechanics, from current documentation:
- Turnitin shows no score between 1 and 19%. The indicator displays an asterisk, no percentage and no highlights, because Turnitin found more false positives in that range.
- A submission needs 300 to 30,000 words of prose, in a .docx, .pdf, .txt or .rtf file under 100 MB, in English, Spanish or Japanese.
- Nothing is retroactive. Turnitin updates the model regularly, and says it folded an ensemble into a single model in July 2026, but no update changes an existing report. It has to be resubmitted, and can come back with a different number.
Can Turnitin detect ChatGPT?
Often, yes, when model output is pasted in unedited. Turnitin publishes the models its English detector is trained to recognise: as of August 2026, GPT-5 and its variants, Gemini 3, Claude, Llama, Mistral, DeepSeek and Grok, plus tools built on them. It says it also catches AI text since pushed through a paraphraser or bypasser, in English submissions only. It cannot establish authorship, though: it returns a likelihood that a passage is machine-generated, not evidence about who sat at the keyboard. Text a person has genuinely rewritten, rather than run through another tool, is the hardest case for any detector, and Turnitin says it would rather miss AI writing than raise its false positive rate. It is also the case that stops being AI slop at all, since that word is about absent effort rather than origin.
Does every school turn it on?
No. AI writing detection is a licensed feature, not a default. Turnitin says it comes with Turnitin Originality or as an iThenticate 2.0 add-on, administrators can enable or disable it, and where it is off, submissions are not processed for AI detection at all.
The University of Waterloo discontinued it in September 2025, citing unreliability, bias against students whose first language is not English, and its own testing in which human-written text came back flagged as 100% AI. Vanderbilt disabled it in August 2023. Curtin University said in September 2025 it would switch the feature off from January 2026 while keeping text matching. Whether your work gets checked this way depends on your institution, and the only way to know is to ask.
Can students see their own Turnitin AI score?
No. Turnitin says students cannot see the AI writing indicator or the report, though an instructor can download it as a PDF and share it, which is worth asking for. Nor is there a public turnitin.com AI detector you can paste an essay into: Turnitin sells to institutions and publishers, not individuals, and the indicator only appears on work submitted through an institutional account. Anything advertising itself as a free Turnitin AI checker is not running Turnitin's model and cannot tell you what a report would say.
False positives, and who they land on
Turnitin says it aims to keep the false positive rate, meaning fully human writing marked as AI, under 1% for documents with over 20% AI writing, roughly one human document in a hundred, and that before each model update or release it tests against over 700,000 academic papers written before ChatGPT existed.
A document rate is not a sentence rate. In a June 2023 post Turnitin put the sentence-level rate near 4%, and said those sentences tend to sit beside genuine AI writing. Its current FAQ publishes the document figure only.
Turnitin also documents where its own detector goes wrong, the most useful thing on any of its pages:
- Writing with little structural variation, repetitive text, and paraphrase that develops no new idea, the shape of formulaic academic prose.
- Short submissions. Under a few hundred words there is no room to overlap segments, so the prediction goes near all or nothing and mixed writing can come back as entirely AI.
- Generic openings and closings: Turnitin changed its detection logic in May 2023 after finding more false positives in a document's first and last few sentences.
- Grammarly, in part: Turnitin says the detector does not target spelling, grammar and punctuation fixes, but content from Grammarly's generative features, including drafting, paraphrasing and summarising, will likely be flagged.
Who this lands on is still argued, and most independent research covers detectors in general, not this one. The 2023 Stanford study in Patterns, the most cited work on detectors misreading writers whose first language is not English, did not test Turnitin; Turnitin says it deliberately included second-language learners and under-represented groups in its training sample to reduce that bias. A 2023 test of fourteen tools did include Turnitin and found the detectors of the time neither accurate nor reliable, obfuscation making them worse. Both predate several rebuilds of the model, so treat them as grounds for caution, not current measurements. More of that evidence, and what universities decided after weighing it, is on our page about the AI detectors schools and colleges use.
What to do about a flag
If a Turnitin AI percentage has been put to you, ask four questions.
- What was the number? Below 20% there is only an asterisk, so "Turnitin flagged it" can mean less than it sounds.
- What else is this resting on? Turnitin does not determine misconduct, and a percentage on its own is not a case.
- What does the policy say? Institutional policy sets the consequence, not Turnitin, and institutions differ: some act on a flag, others switch the feature off. Turnitin's own advice to instructors is to use a highlighted passage to start a conversation, not draw a conclusion.
- Can I see the report? Not by default, but an instructor can export it.
Then gather process evidence, the only thing you hold that speaks to authorship: version history in Google Docs or Word, dated drafts, notes, outlines, the browser history of what you read. Turnitin publishes a guide for students on handling false positives that points the same way. None of it guarantees an outcome, and anyone who says it does is selling something.
A second reading is worth having before that conversation, and that is what this site does. Paste the passage and you get a likelihood plus the parts of the text behind it, so you see what a detector reacts to, not just a number. It judges where the words came from, not whether they are any good. It cannot see your Turnitin report, will not overturn one, and is not proof of anything. What it does is show its working.
Last updated 2026-08-17.