Turnitin's AI writing indicator is a panel inside the report an instructor opens after a paper is submitted, at institutions that license Turnitin. It shows a percentage meant to represent how much of the qualifying prose in a submission its model scored as likely AI-generated. Two things about it surprise most people who search for it. Students cannot run it themselves, because Turnitin sells to institutions rather than to individuals and the indicator sits on the instructor's side of the assignment. And the percentage is not the similarity score, which is the older number people picture when they hear the name. We have no affiliation with Turnitin, and nothing here is a claim about what its model will return on your text.
Institutions license Turnitin and wire it into their assignment workflow, usually through the learning management system, so handing in a paper through Canvas, Moodle or Blackboard can mean submitting it to Turnitin at the same time. The instructor then opens a report on that submission, where the similarity view has been a fixture for years. The AI writing indicator is a separate panel with its own percentage and its own view of the document, highlighting the passages the model scored as likely machine-written.
The model reads the text and nothing else. It does not see your file history, your keystrokes, the hours you worked, or which program you wrote in. What it returns is a statement about how the prose reads, not an observation of how the document was made, so it is not evidence of authorship. It is a statement that some sentences resemble the patterns the model associates with machine-generated writing.
Whether the indicator appears at all is a local setting. Administrators can switch it on or off, and institutions have gone both ways. Turnitin's own guidance for instructors presents the figure as an indicator rather than a determination of misconduct, and points toward a conversation with the student. Read the current documentation before quoting it, because the vendor's own framing carries more weight in a meeting than a paraphrase of it.
There is no consumer plan. Access runs through an institution's licence, and inside that licence the AI writing indicator is an instructor-facing view. Students who submit to an assignment do not see it. Even where a school opens a draft folder so students can check their own similarity report before the deadline, the AI figure is not part of what the student is shown.
So there is no pre-check. You cannot see the number before you submit, you usually cannot see it afterwards unless your instructor shows you, and you cannot re-run it on a revised paragraph. Advice that tells you to keep testing until the score drops describes a loop students do not have access to.
A check bought from a third-party site does not answer the question either. Whatever number comes back came from a different model on a different upload, and it stands in no relationship to the report inside your institution's account. If you have been flagged, ask to see that report. Read the terms of anything you paste an unsubmitted draft into, because services differ on what they store.
Similarity is a matching operation. Turnitin compares your text against its repositories of previously submitted student papers, indexed web pages and publisher content, then reports the share of your submission that matches something it holds. Every match has a source behind it that you can click through and read. A high similarity score is often nothing: quoted material, a reference list and standard phrasing in your field all match.
The AI writing indicator has no sources behind it. There is nothing to click through to, because nothing was matched. It is a model's estimate about style. A paper can come back with almost no similarity and a high AI figure, or the reverse.
The indicator works on long-form prose, and a submission needs enough of it to qualify. Turnitin documents a minimum length, and its documentation is where to check the current figure. Short answer sets, bullet lists, code, tables and poetry fall outside what the model scores. A submission below the floor comes back with no score, which is not a clean bill of health. It means the tool was out of scope, not that it looked and approved.
The percentage is a proportion of text, not a level of confidence. The model scores sentences inside the qualifying prose and reports the share it marked as likely AI-generated. A high figure is not the tool announcing how sure it is. It is saying roughly that share of your sentences crossed its line, and one classifier's line is not a fact about who wrote them.
Language coverage is limited and has been extended over time, so a paper written in a language the model does not cover comes back with nothing. Turnitin has also described reporting AI-paraphrased text separately from the main figure, so if you are quoted a number, ask which number it is. Both move as the product changes, so check the documentation rather than a forum post.
The strongest published finding about detection and ordinary writers comes from Stanford HAI. Liang and colleagues ran TOEFL essays written by non-native English speakers through seven AI detectors and found that 61% were falsely flagged as AI-generated. Every one of those essays had a human author. That study did not test Turnitin and says nothing about Turnitin's rate. It is evidence about the class of tool, and about which writers absorb the mistakes.
The writing that pushes a detector's score up tends to be even, plain, correctly punctuated and organised to a template. That describes second-language writers, people taught a rigid five-paragraph structure, people who use grammar checkers or dictation, and plenty of careful writers besides. If you write that way, a high number may be describing your habits rather than your sources. That is the argument to make in the room, with the study cited.
The percentage settles nothing on its own, so find out what is being claimed and on what basis. Keep the questions short and put them in writing, because a written exchange creates its own record.
Then preserve the record before it degrades. In Google Docs, File then Version history keeps timestamped revisions. In Word, check AutoSave and the OneDrive or SharePoint version list. Do not copy the file into a fresh document, because that is how people destroy the best evidence they have. Gather outlines, earlier drafts, notes and any message you sent about the work, since those carry timestamps you did not control.
Write a short factual account of how you produced the piece: when you started, what you read, what you cut, and which tools you used, including a spell checker or a translation app. A specific account you can repeat consistently does more than a general denial. The longer version, with those questions written out as sentences you can send, is on our page for writers who have been accused.
Our checker is a separate tool with its own model and its own scoring, and it answers a narrower question: how does this text read to this checker. It returns a 0 to 100 score and one of three readings, "Reads as AI-generated", "Borderline" or "Reads as human-written", in seconds, on anything from 65 words up. One run costs one credit whatever the length. The free tier is 10 credits a day per network, with no account, no login, no card and no email address. It cannot tell you what Turnitin will report on the same text, and we make no claim about that.
For the question most people actually have, which is which sentences read as machine-written, the registry of AI writing tells is more direct. It runs in your browser, free and unlimited, and highlights the exact phrases a detector keys on, so you can see what is driving the impression. The proofreader, word counter, readability checker, AI vocabulary checker, plagiarism and originality check and signed drafting record sit next to it, also free and unlimited.
On policy, once and briefly. Rewriting a draft you were permitted to generate is ordinary editing. Where AI is not permitted, that rule covers the draft however it is edited. Read the policy that applies to your course and follow it.
No. Turnitin licenses to institutions rather than to individuals, and the AI writing indicator is a view on the instructor's side of an assignment. Students do not see it, including in draft folders where the similarity report is visible. If you want to know what your instructor saw, the only route is to ask them for the report.
No. They are separate figures in the same report. Similarity measures matched text and links to the sources it matched. The AI indicator is a model's estimate of how much qualifying prose reads as machine-written, with no sources behind it. A paper can score low on one and high on the other.
It refers to the share of the qualifying prose that the model scored as likely AI-generated. It is not a confidence level, and it is not a proportion of the paper copied from anywhere. It is a proportion of sentences, produced by a classifier that can be wrong, and Turnitin's own guidance presents it as an indicator rather than a determination of misconduct.
The usual reasons are scope. The indicator needs long-form prose above a documented minimum length, so short answers, bullet lists, code, tables and poetry return nothing. An unsupported language returns nothing. Your institution may also have the indicator switched off. No score is not the same as a score of zero.
Nobody can tell you what a given detector will output on a given document. What the published research describes is a pattern: even, correctly punctuated, plainly structured prose draws higher scores across detectors generally, and that is what heavy editing and translation tend to produce. If you are asked, say plainly which tools you used. It is a normal answer, and it is the answer a careful writer gives.
We do not test against Turnitin and we make no claim about what it returns before or after anything. Our humanizer rewrites every sentence in one click, from 25 words up to 12,000 characters or roughly 2,000 words per run, at 1 credit per 50 words. What any third-party detector does with the output is not something we measure, so treat a promise of a specific score as a guess.