Can Turnitin Detect AI? Accuracy, False Positives, and Score Meaning

Yes. Turnitin can flag passages it predicts were generated by AI or generated by AI and then paraphrased. But an AI Writing Score is not proof that a student cheated, and it is not the same as Turnitin's Similarity Score.

Here’s a product demo of TurnItIn:

Turnitin's own guidance says its model can misidentify human-written, AI-generated, and AI-paraphrased text. The company instructs educators not to use the report as the sole basis for adverse action against a student. A report must be interpreted alongside the assignment rules, the student's writing process, other evidence, and human judgment.

Question	Evidence-based answer
Can Turnitin detect AI writing?	It can identify qualifying prose that its model predicts was AI-generated, AI-paraphrased, or modified by certain AI bypasser tools.
Is the result definitive?	No. Turnitin describes the report as one indicator that requires further review and human judgment.
What does `*%` mean?	The model detected some AI-like text, but the result was below 20%, where false positives are more common. Turnitin hides the exact number and highlights.
Is the AI score a plagiarism score?	No. The AI Writing Score and Similarity Score are separate measurements.
Does a 40% score mean Turnitin is 40% certain?	No. It means the model classified approximately 40% of the qualifying prose as likely AI-written or AI-altered.
Does 0% prove that no AI was used?	No. It only means the model did not identify qualifying text as likely AI-written in that submission.
Can every file receive an AI score?	No. The file must meet Turnitin's language, length, format, and prose requirements.
Do all universities treat the report the same way?	No. Access, interpretation, investigation, and permitted AI use vary by institution, course, and assignment.

AI Writing Score versus Similarity Score

Students frequently confuse these two reports, but they answer different questions.

Report	What it measures	What it does not prove
Similarity Score	The percentage of text that matches material in Turnitin's comparison databases	That the matching text is plagiarism or academic misconduct
AI Writing Score	The percentage of qualifying prose the model predicts was likely AI-generated or AI-generated and then altered	Who wrote the work, which exact tool was used, or whether misconduct occurred

A quotation may increase a Similarity Score even when it is properly cited. Conversely, AI-generated wording may have little similarity to existing sources. That is why a paper can have a low Similarity Score and a higher AI Writing Score or the reverse.

Turnitin places the AI report within its review workflow, but the percentages are independent. Educators must interpret both reports rather than treating either one as an automatic verdict.

How Turnitin's AI Writing Report works

Turnitin says it breaks a submission into overlapping text segments so that sentences can be evaluated in context. Its model scores the text and estimates how much of the qualifying prose is likely to have originated from generative AI.

The report can divide detected text into categories:

Likely AI-generated text, including text the model predicts may have been modified using a bypasser tool, is highlighted in cyan.
Likely AI-generated and AI-paraphrased text is highlighted in purple.

Turnitin introduced AI-paraphrasing detection in 2024 and added detection aimed at AI bypasser tools in 2025. Its product documentation says AI paraphrase and bypasser detection are currently available for English submissions, while general AI writing detection supports English, Spanish, and Japanese.

These labels describe model predictions. They do not establish which application produced the text or whether the student's use was allowed.

What counts as qualifying text?

Turnitin's detector is designed mainly for long-form prose: complete sentences in paragraphs that form an essay, article, or dissertation.

For an AI Writing Report to be generated, Turnitin currently requires:

A file smaller than 100 MB
At least 300 words of long-form prose
No more than 30,000 words
A supported language: English, Spanish, or Japanese
A supported file type: docx, .pdf, .txt, or .rtf

Turnitin says the model does not reliably evaluate non-prose or unconventional formats such as:

Poetry
Scripts
Computer code
Bullet-point lists
Tables
Annotated bibliographies

A mixed-format document may therefore show a percentage that applies only to part of the visible submission. The score is based on qualifying text, not necessarily every word in the file.

What does a Turnitin AI score mean?

0%

A 0% result means Turnitin did not classify any qualifying text as likely AI-written or AI-altered.

It does not prove that:

No AI tool was used
Every sentence was written without assistance
The work complies with the assignment policy
A different detector would return the same result

*%

An asterisk means the detector found a result above 0% but below 20%.

Turnitin stopped displaying exact scores and highlights in this range because its testing found a higher incidence of false positives. A *% result is deliberately presented as less reliable.

It should not be interpreted as:

A hidden 20% score
Proof of some AI use
An official "safe" result
A reason to rewrite text to manipulate a detector

20% or higher

At 20% or above, Turnitin displays an exact percentage and highlights the passages contributing to the score.

The percentage represents the estimated share of qualifying prose predicted to be AI-written or AI-altered. It is not a probability of guilt.

For example:

A 40% AI Writing Score does not mean Turnitin is 40% confident that misconduct occurred.

It means the model classified approximately 40% of the qualifying prose in the submission as likely belonging to its AI-detection categories.

Gray indicator or processing error

A paper may receive no score because:

It did not meet the file requirements
AI detection was not enabled when it was submitted
The document could not be processed
The report encountered a technical error

"No score" is different from 0%.

Is there a safe Turnitin AI percentage?

No universal safe percentage exists.

Turnitin does not decide whether academic misconduct occurred, and it does not publish a universal disciplinary threshold for schools. A university, department, instructor, or assessment may apply different rules.

Even a high score should be reviewed in context. A low or 0% score does not certify acceptable AI use.

The questions that matter are:

What did the assignment permit?
Was AI use disclosed where required?
Can the student explain the argument, sources, and writing process?
Is there additional evidence beyond the detector score?
What procedure does the institution require?

How accurate is Turnitin's AI detector?

There is no single accuracy percentage that applies to every student, language, discipline, AI model, document length, and type of AI assistance.

What Turnitin reports

Turnitin reports a false-positive rate below 1% for documents containing more than 20% AI-generated content under its testing conditions. It also explicitly acknowledges that false positives and false negatives can occur and says the report should not be used by itself to penalize a student.

The scope matters. A vendor-reported false-positive rate is not the same as overall accuracy across:

Human writing
Fully AI-generated writing
AI-assisted human writing
Hybrid documents
AI-paraphrased text
Different academic disciplines
Different languages
Newer generative models

What independent studies report

Independent results vary considerably.

A 2026 peer-reviewed study by Hadra, Cambridge, and Mesbah evaluated Turnitin and Originality using 192 human, AI-generated, and hybrid texts. It reported overall accuracy of 0.61 for Turnitin and found that hybrid writing was particularly difficult to classify. Performance also changed with text length and genre.

Another 2026 peer-reviewed study tested Turnitin, GPTZero, Copyleaks, and Pangram using 160 documents with known human, AI, hybrid, and "humanised" ground truth. In that dataset, Turnitin substantially underestimated AI content, especially in fully AI-generated documents produced using newer workflows. The authors also noted that detector performance depends on the test design and detector version.

Earlier research by Weber-Wulff and colleagues found that AI-text detectors were not consistently accurate or reliable and became less effective after text modification. That study is useful historical evidence, although Turnitin has updated its model several times since the research was conducted.

These studies do not establish that Turnitin always performs poorly. They demonstrate why a single marketing number should not be treated as a universal measure of reliability.

Detector versions change

Turnitin updated its AI writing model in February 2026 to improve recall while maintaining a low false-positive rate. It states that previously generated reports are not recalculated automatically. A paper must be resubmitted to receive a score from the updated model.

That means:

The same document can receive a different result after a model update.
An accuracy study should identify the detector version and test date.
Old screenshots and percentages may not represent the current model.

Can Turnitin produce false positives?

Yes. A false positive occurs when human-written text is classified as likely AI-written.

Turnitin has acknowledged that formulaic introductions and conclusions previously produced a higher incidence of false positives and has adjusted its model. It also suppresses exact results below 20% because that range is less reliable.

False positives may be more concerning in writing that is:

Highly formulaic
Technically repetitive
Built around standard disciplinary phrases
Written to a rigid template
Short or lacking enough qualifying prose

However, no writing style automatically causes a false positive, and students should not deliberately make their writing worse or less clear to influence a detector.

Can Turnitin miss AI-generated text?

Yes. A false negative occurs when AI-generated or AI-assisted writing is not classified as AI.

Possible reasons include:

The text does not meet the qualifying-prose requirements.
The document contains a mix of human and AI-authored material.
The model does not recognize the output pattern.
The writing was heavily revised.
The detector version was built before a newer generation method.
The submission is in a format or language with more limited support.

A missed detection does not make prohibited AI use acceptable. Academic integrity depends on following the assignment rules, not on whether software flags the work.

What do universities do with a high score?

Policies differ substantially.

University of Melbourne

The University of Melbourne tells students that an AI report is a prompt for further investigation, not sufficient evidence on its own. A student may be asked to explain the argument, sources, and writing process or provide drafts and notes. Its staff guidance requires additional evidence before a misconduct allegation is made.

Vanderbilt University

Vanderbilt disabled Turnitin's AI detector in 2023 after raising concerns about reliability, transparency, privacy, and the effect of false positives at institutional scale. This is an example of an institution choosing not to use the feature rather than evidence that all universities have done the same.

University of Cambridge

Cambridge explains that Turnitin use varies by faculty, course, and assessment. Reports are reviewed by academic staff, and possible concerns require further investigation under university procedures.

The practical lesson is simple:

Your institution's published policy and your assignment instructions matter more than general advice found online.

Can students see the AI Writing Report?

Not always.

Turnitin's AI Writing Report is primarily designed for educators. Student access depends on the product, institutional settings, and assignment configuration.

For example, the University of Melbourne says its students do not automatically see the AI indicator or highlighted report, although an instructor may share a PDF when discussing concerns. Turnitin Clarity has separate Writing Report access settings that instructors can enable for students.

Do not assume that a student-facing Similarity Report automatically includes the AI Writing Report.

What should you do if your work is questioned?

Being asked to discuss a paper does not necessarily mean you have been found responsible for misconduct.

1. Read the exact assignment policy

Identify:

Whether AI was prohibited, permitted, or permitted with disclosure
Which forms of assistance were allowed
Whether citations or an AI-use statement were required
Which institutional procedure applies

2. Ask to review the concern

Politely ask:

Which passages raised concern?
Was the concern based only on the AI Writing Report?
Can you review the report with the instructor?
What policy or assessment rule is being applied?

3. Gather evidence of your writing process

Useful records can include:

The assignment prompt
Brainstorming notes
Search history related to research
Source PDFs or reading notes
An outline
Early drafts
Document version history
Instructor or peer feedback
Revision notes
Citation records
Any required AI-use disclosure
A timeline of how the work developed

These records do not automatically determine an outcome, but they can help explain authorship and process.

4. Be ready to explain the paper

You may be asked to explain:

The thesis
Why you selected particular sources
How the evidence supports each claim
How your argument changed during drafting
The meaning of important concepts
Which tools you used and for what purpose

5. Correct genuine problems honestly

If you used an AI tool in a way that was permitted but failed to disclose it, explain what happened and follow the instructor's guidance.

If a source, quotation, or citation is inaccurate, acknowledge and correct it. Do not fabricate drafts or evidence.

6. Follow the formal review process

When an informal conversation does not resolve the issue, use your institution's published academic-integrity or appeal procedure. Keep communication factual and organized.

Proof-of-writing-process checklist

Save these materials while working rather than reconstructing them after a concern is raised.

Assignment instructions and rubric
Topic ideas or brainstorming
Research notes
Copies or links to sources
Outline
First draft
Version history
Feedback received
Revision notes
Citation library
Required AI disclosure
Final pre-submission copy

How to use AI tools responsibly

The responsible use of AI depends on the rules for the specific course and assignment.

Check the policy before using a tool

Do not assume that an AI feature is allowed because it is available in common writing software. Policies may distinguish among:

Brainstorming
Outlining
Grammar correction
Translation
Paraphrasing
Generating sentences or paragraphs
Summarizing sources
Producing citations
Writing code
Analyzing data

Ask the instructor when the boundary is unclear.

Keep the intellectual work yours

Your submission should reflect your understanding of the topic. You should be able to explain and defend the argument without relying on a generated answer.

Verify every factual claim and citation

Generative AI can invent:

Sources
Authors
Page numbers
Quotations
Statistics
Laws
Study findings

Read the original source before using it.

Disclose use when required

A disclosure can identify:

The tool
The purpose
The part of the process where it was used
What you changed or verified
Whether generated material appears in the final submission

Follow the required format rather than inventing your own when the institution provides one.

Do not use "humanizer" tools to bypass detection

Changing text to manipulate an AI detector does not resolve an authorship or policy problem. It may also introduce factual errors, awkward wording, or additional integrity concerns.

FAQs

Can Turnitin detect ChatGPT?

Turnitin is designed to identify text it predicts was produced by generative AI, but it cannot guarantee detection of every ChatGPT output. Results depend on the model version, document type, language, length, and how the text was used or revised.

Can Turnitin detect GPT-5, Gemini, or Claude?

Turnitin updates its detector as generative models change, but it does not provide a universal guarantee for every output from every model. Its current product updates mention support improvements for newer models in specific language models, but detector performance still depends on the submission and detector version.

Can Turnitin detect AI paraphrasing?

Turnitin can label text it predicts was AI-generated and then AI-paraphrased. Its current documentation says AI paraphrase and bypasser detection are available for English submissions. A label remains a model prediction, not proof of the tool used.

Can Turnitin detect Grammarly?

There is no reliable yes-or-no answer for every Grammarly feature. Basic spelling or grammar correction is different from generating or extensively rewriting text. The important question is whether the specific feature was permitted and whether disclosure was required.

Does a high AI score prove cheating?

No. Turnitin says the AI report should not be the sole basis for adverse action. An educator should consider the assignment policy, the student's process, the highlighted text, other evidence, and a conversation with the student.

Does *% mean less than 20% AI?

It means the model produced a result above 0% but below 20%. Turnitin does not reveal the exact number or highlights because that range is less reliable.

Is 20% AI acceptable?

Turnitin does not define 20% as acceptable or unacceptable. The percentage is not a universal misconduct threshold. The course or institution determines what use is allowed.

Can I check my own Turnitin AI score before submitting?

Usually not through Turnitin unless your institution provides access. Avoid uploading unpublished work to random third-party detectors, because their accuracy and data-handling practices may be unclear.

What if I wrote the paper myself but the score is high?

Gather your drafts, notes, sources, and version history. Ask to review the report and explain how you developed the argument. Follow the institution's formal process if the concern continues.

Key takeaways

Turnitin can flag qualifying prose it predicts was AI-generated or AI-paraphrased.
The AI Writing Score is separate from the Similarity Score.
The percentage describes qualifying text, not certainty that misconduct occurred.
*% represents a result below 20%, where Turnitin considers the output less reliable.
Turnitin can produce false positives and false negatives.
Independent studies show that performance varies by genre, length, model, and mixed authorship.
Universities use different policies and procedures.
The strongest protection for honest work is a clear writing process supported by drafts, notes, sources, and version history.
Students should follow the assignment's AI policy rather than trying to influence a detector.

About this guide

Author: Daniel Felix

Last evidence review: July 17, 2026

Evidence used: Official Turnitin documentation, university academic-integrity guidance, and peer-reviewed detector research