Turnitin's AI writing indicator estimates the share of qualifying prose that its model considers likely to be AI-generated or AI-modified. It appears inside eligible institutional Turnitin reports and is separate from the Similarity score. A 40% AI indicator does not mean the paper is 40% plagiarized, and it is not a statement that Turnitin identified a particular chatbot or proved misconduct.
The most important rule comes from Turnitin itself: the AI score should not be used as the sole basis for adverse action against a student. The report can direct an educator to passages worth reviewing. A fair decision still requires the assignment policy, the student's explanation, drafts, notes, sources, version history and the educator's knowledge of the work.
Turnitin AI detection fact | Current position | Why it matters |
|---|---|---|
Where it appears | Inside supported institutional Turnitin workflows | It is not a public student self-checker |
What it scores | Qualifying long-form prose | The denominator may exclude lists, code and references |
Relationship to Similarity | Independent score | AI likelihood and source matching answer different questions |
Low-score handling | 1% to 19% is shown as an asterisk | Turnitin withholds the number to reduce false-positive misuse |
Decision role | Investigative signal | The report is not a misconduct verdict |
What Turnitin AI detection is
Turnitin AI detection is a feature attached to certain institutional licenses and report experiences. Availability depends on the product, add-on and administrator settings. Instructors generally reach the AI Writing Report from the Similarity Report. Students do not receive a consumer version that lets them upload a paper and see the same institutional indicator.
The indicator reports a percentage of qualifying text, not a percentage of the entire uploaded file. Turnitin defines qualifying text as prose sentences in a long-form writing format. Material the model does not evaluate can still be present in the submission. This distinction explains why the percentage and visible highlights may not correspond to a reader's rough count of every word on the page.
How the Turnitin model works
Turnitin describes its current detector as a proprietary transformer-based deep-learning model trained for academic writing. It processes a submission in overlapping text segments, assigns predictions within those segments and aggregates them into sentence and document-level results. Overlap gives the model context, but a short submission provides fewer stable segments and can produce a blunter prediction.
This matters because many generic explanations say all detectors simply check perplexity and burstiness. Those concepts can help explain AI-text classification, but Turnitin does not present its report as a transparent formula built from two user-readable features. Its classifier learns many patterns from training examples. The practical way to judge it is through current validation on relevant academic text, not through a simplified description of one signal.
How to read the AI writing indicator
Indicator | Meaning | Recommended response |
|---|---|---|
0% or a numeric score from 20% to 100% | The document was processed and the model returned a reportable estimate | Open the report and review qualifying passages in context |
*% | Some AI-like text was detected between 1% and 19%, but the percentage is withheld | Do not infer a hidden exact score or treat the asterisk as proof |
Dash | No score is available, often because requirements were not met | Check language, prose length, file type and processing status |
Error state | The report could not be completed | Resubmit when appropriate or contact institutional support |
The asterisk is a deliberate safeguard. Turnitin says low percentages have a higher incidence of false positives, so reports between 1% and 19% do not receive a number or highlights. An educator should not convert that symbol into a disciplinary threshold. It means the product declined to provide a precise result in that range.
A reportable percentage is still an estimate. It does not identify which account generated the text, whether the student had permission to use AI, or how much human revision occurred. The blue highlights show where the model found relevant patterns. They are starting points for review, not a reconstruction of the writing process.
Turnitin AI score and Similarity score are separate
Report | Question answered | Evidence shown | Common mistake |
|---|---|---|---|
AI Writing Report | Does qualifying prose resemble AI-generated or AI-modified writing? | Predicted percentage and highlighted passages when reportable | Treating probability as proof of misconduct |
Similarity Report | Does language match indexed sources or submissions? | Matched passages and source records | Treating every match as plagiarism |
Draft and process evidence | How did the student develop the work? | Notes, versions, citations and explanation | Ignoring stronger contextual evidence |
A paper can have a low Similarity score and a high AI indicator because generated text may not match an indexed source. It can also have a high Similarity score and a low AI indicator because human-written text contains quotations, templates or copied language. Educators should inspect each report for the question it was designed to answer.
For broader product comparisons, see Best AI Detectors in 2026 and Best Plagiarism Checkers.
Current file and text requirements
Requirement | Turnitin documentation | Review note |
|---|---|---|
Prose length | At least 300 words and no more than 30,000 words | A document may be longer overall if little of it qualifies as prose |
Supported languages | English, Spanish, Japanese and Modern Standard Arabic in current report guidance | Language models and feature coverage may differ |
Accepted file types | .docx, .pdf, .txt and .rtf | Scanned image-only files will not provide usable text |
File size | Under 100 MB | Institutional upload rules may add other limits |
Text form | Long-form prose sentences | Code, lists, tables and references may be excluded or unsupported |
Turnitin documentation is updated as language and model coverage changes. Its current Using the AI Writing Report guidance includes Arabic alongside English, Spanish and Japanese, while some older file-requirement snippets may still list only three languages. Institutions should rely on the newest report guidance and release notes, then confirm behavior in their own account.
The 300-word minimum is not a promise that every 300-word document will be equally reliable. It is an eligibility floor. More qualifying prose generally gives a classifier more evidence, while mixed, formulaic or highly edited writing remains difficult even at greater length.
What Turnitin says about accuracy
Turnitin states that it aims to keep the false-positive rate below 1% for documents with more than 20% AI writing. That is a vendor claim tied to its validation data, threshold and model version. It does not mean every highlighted sentence is 99% certain, and it does not establish the false-negative rate for every current generator or editing method.
The product is intentionally conservative in some conditions. Withholding exact results below 20% reduces the chance that weak signals are overinterpreted, but conservative thresholds can miss AI-written passages. A detector can have a low false-positive rate and still have a meaningful false-negative rate. Institutions need both measurements before deciding whether a tool is fit for their assignments.
Independent evidence has changed over time. A 2023 comparison reported strong Turnitin performance on its dataset. A peer-reviewed 2026 study of four detectors found substantial variation across fully human, fully AI, hybrid and humanized academic papers, including serious missed-detection problems for Turnitin on that test set. The AI text in the study predated several later model updates, so the finding should prompt current local validation rather than be treated as a permanent score for every Turnitin version.
Independent study: Who wrote this? Evaluating the reliability of AI detection tools in higher education.
A repeatable institutional test before relying on the score
Test group | What to include | Metric to record |
|---|---|---|
Verified human work | Past student papers across writers, subjects and proficiency levels | False-positive rate overall and by group |
Current AI output | Several current generators using real assignment prompts | Detection rate by model and length |
Hybrid papers | Known AI passages inserted at several proportions | Boundary accuracy and document score |
Human-revised AI | Light, moderate and substantial documented revisions | Change in detection after editing |
Supported languages | Authentic work in every language the institution uses | Performance by language and genre |
Repeated version test | The same set after major Turnitin updates | Model drift and threshold changes |
The test should be preregistered before reviewers see the labels. Record the date, Turnitin product, model or release state, assignment type and exact interpretation rule. Report false positives and false negatives separately. A balanced test set is useful for model comparison, but the institution should also estimate how common prohibited AI use is in practice because that affects how many flagged papers are likely to be true cases.
How educators should review a high Turnitin AI score
- Open the AI Writing Report and identify which qualifying passages were highlighted.
- Check that the submission met the language, format and prose-length requirements.
- Read the assignment policy and distinguish permitted assistance from prohibited generation.
- Compare the passages with drafts, notes, sources, version history and earlier work.
- Ask the student to explain the argument, evidence, vocabulary and revision process.
- Consider language background, accessibility tools and legitimate editing support.
- Document the evidence and provide the same review and appeal process to every student.
The conversation should begin with the work, not an accusation. Ask open questions about how the student selected sources, developed a claim and revised a specific paragraph. A student who can explain the decisions may provide process evidence that the classifier cannot see. A student who used prohibited assistance should still be evaluated under a written policy and consistent procedure rather than an improvised score cutoff.
The broader classroom workflow is covered in Best AI Tools for Teachers. Educators who want a separate, inspectable second analysis can review the Winston AI content detector, but disagreement between tools should lead to more context, not a detector vote.
What students should know
Students usually cannot access the same institutional Turnitin AI indicator before submission. Websites claiming to provide a public Turnitin AI score should not be assumed to be Turnitin. The best preparation is to follow the assignment policy, keep notes and drafts, cite assistance when required and preserve version history that shows how the work developed.
If a paper is flagged, ask to see the report and the passages under review. Explain the writing process with concrete evidence. Identify any tools used for grammar, translation, accessibility, research or generation, and describe how their output entered the final submission. A clear account is more useful than trying to prove innocence with screenshots from several unrelated detectors.
Turnitin's strengths and limits
Turnitin's main strength is workflow. The indicator sits beside source-matching tools already used by institutions, supports passage-level review and publishes unusually specific cautions about low scores and adverse action. The 20% reporting rule helps prevent weak low-band estimates from appearing more precise than they are.
Its limits are equally important. Students lack an equivalent self-check, the classifier cannot observe the writing process, model performance changes with new generators and editing, and institutional access can encourage users to confuse availability with certainty. A responsible policy treats Turnitin as one source of information and builds the final judgment around process evidence and human review.
Frequently asked questions
Can Turnitin detect ChatGPT?
Turnitin says its current models cover writing from major language-model families, including current OpenAI systems. Coverage does not guarantee detection of every passage, especially after editing or below reporting thresholds.
What does *% mean in Turnitin?
It means the detector found a low level of AI-like text between 1% and 19%, but Turnitin withheld the exact percentage and highlights because false positives are more likely in that range.
Is a Turnitin AI score the same as a Similarity score?
No. The AI indicator estimates AI-like prose. The Similarity score measures matching language against indexed sources. They are independent reports.
Can students check their Turnitin AI score?
The institutional AI Writing Report is not a public student self-checker. Access depends on the institution and report permissions.
Is a 100% Turnitin AI score proof?
No. It is a model estimate about qualifying prose and should be reviewed with the assignment, passages, drafts, sources, version history and student explanation.
Does Turnitin detect Grammarly or paraphrasing?
Turnitin says ordinary spelling, grammar and punctuation corrections are not its target. Generative drafting, summarizing, paraphrasing and bypasser-modified text may be included in AI detection.
What should an educator do after a high score?
Inspect the highlighted text, verify requirements, review process evidence, speak with the student and follow a documented institutional policy. Do not impose an automatic penalty from the score alone.
Sources and methodology notes
This guide was checked on September 16, 2026. Current capability statements come from Turnitin's AI writing detection capabilities FAQ, Using the AI Writing Report, File requirements for an AI Writing Report, AI model release notes and educator review guidance. Turnitin is named without outbound product links under the site's editorial policy.
Vendor claims are labeled as such and are not compared directly with other tools unless the dataset, labels, thresholds and versions match. Independent findings are presented with date and version caveats because detector and generator behavior changes.
Related reading: How AI Detectors Work, Best AI Detectors in 2026, and Best AI Humanizer.