Turnitin AI Percentage: What the Number Measures

The Turnitin AI detection percentage is the share of qualifying prose the model identifies as likely AI-generated. A 50% indicator does not mean a 50% chance that the author used AI. It also does not necessarily describe half of every word in the submitted file.

The distinction comes directly from Turnitin’s AI Writing Report guide. This page explains the denominator and display states. If you need to obtain the report or prepare a response, follow our report access and evidence workflow.

What counts as qualifying text?

Turnitin describes qualifying text as prose sentences in a long-form writing format: paragraphs in an essay, dissertation or article. It says the model does not reliably detect generated text in non-prose such as poetry, scripts and code, or in short-form and unconventional material such as bullet points, tables and annotated bibliographies.

That changes how you read a mixed document. A report might contain a large data table, a bibliography and several pages of prose. The AI percentage is not a measure of every element equally. Do not multiply the percentage by the file’s total word count and present the result as an exact number of AI-written words.

Question	Correct interpretation	Common misreading
What is the denominator?	Qualifying long-form prose	All content in the uploaded file
What is being summarized?	Text the model identifies as likely generated	A finding of who wrote the work
Does 50% mean uncertainty about the author?	No; it describes a share of qualifying text	“The model is 50% sure I used AI”
Can I count highlighted paragraphs?	Their lengths and eligible content differ	“Three out of six paragraphs is 50%”
Is this the similarity score?	No; the measures are independent	“50% AI means 50% copied”

For the specific arithmetic mistake behind a 50% result, see our worked 50% example.

How to read 0%, *%, and 20–100%

0%. A successfully processed submission had no qualifying text identified as likely AI-generated by the model. This is a result from that model, not certification of human authorship.

*%. For results above 0% and below 20%, the official guide says Turnitin withholds the exact percentage and highlights. Its stated reason is the higher incidence of false positives in that range. Do not infer whether the hidden number is 1%, 10% or 19%.

20–100%. The displayed number gives the amount of qualifying text the model identified as likely AI-generated or generated and then modified in ways covered by its detector. Look at the submission breakdown and highlighted passages, then consider the writing process and applicable policy.

An older report generated before July 8, 2024 may still show a numerical result below 20%. Read its date before comparing it with a newer report. A processing error, loading state or missing report is a separate state, not a zero.

Why the highlights may not match your visual estimate

A visual proportion of highlighted pages measures page layout. The AI percentage concerns qualifying text. A page dominated by a table and a page full of prose are not equivalent denominators, and a paragraph count gives equal weight to passages of different lengths.

Turnitin explicitly notes that mixed writing types can create disparity between the percentage and highlights. That discrepancy does not, by itself, demonstrate a scoring error.

Review the actual submitted file. Were headings, tables or references included? Is the screenshot from the AI report or the Similarity Report? Does the document meet the AI report file requirements? Turnitin documents at least 300 words of long-form prose and no more than 30,000 words.

What the percentage cannot establish

The report does not observe your drafts, research notes or tool use. A high indicator cannot establish intent or misconduct on its own. A low one cannot establish that claims are correct, citations exist or assistance complied with the assignment.

Turnitin’s own guidance says its model may misidentify human-written, generated and modified text, and that its output should not be the sole basis for adverse action. It calls for further scrutiny, human judgment and the institution’s specific academic policies.

This is also why you should not convert a DeAIze score into a Turnitin percentage. They use different instruments, and DeAIze does not offer an official Turnitin readout.

What our separate measurement adds—and does not add

In a dated 16 September 2026 observation, we compared five independent model implementations before and after rewriting six generated drafts. roberta_openai scores rose for four samples, fell for one and stayed the same for one; fast_detectgpt scores fell for four and rose for two. The first-party observation record preserves the counts and limitations.

No binary pass/fail threshold was used for that comparison. Exact model revisions were not stored, and there was no human-written control group. This is not evidence of Turnitin’s accuracy, nor a formula for estimating its result. It is a concrete example of why model-specific score changes cannot be treated as interchangeable.

Our methodology explains the distinction between an internal pattern score and a result actually returned by a third-party detector.

What to do after reading the percentage

Keep the original file and the dated report. Identify what was actually flagged, match the passages to drafts and source notes, and ask what policy and review procedure apply. If the question is about a disputed result, bring that process evidence rather than offering another detector’s number as proof.

If you want a separate reading of an editable draft, try the DeAIze detector with these limits in mind. If you later revise, use the draft editor on a copy and review every factual claim yourself. The aim is a clearer document you can explain, not a predicted Turnitin percentage.