Detect tower · floor

Turnitin AI detection: what the score is

turnitin ai detection is the one nobody chooses. Students meet it because their institution bought it, and instructors meet it because it appeared inside a tool they were already using. That makes it the most consequential detector in education, and the one where knowing what the number means matters most.

Turnitin publishes more about its own limitations than any other company in this field. Two of those published figures are the whole of this floor.

<1%

On a whole document

Turnitin: the document false positive rate — “incorrectly identifying fully human-written text as AI-generated within a document” — is “less than 1% for documents with 20% or more AI writing”.

~4%

On a single sentence

Turnitin: “Our sentence-level false positive rate is around 4%. This means that there is a 4% likelihood that a specific sentence highlighted as AI-written might be human-written.”

Both figures are published by Turnitin itself, read 21 August 2026. They are four times apart, and the smaller one comes with a condition attached — it applies to documents already scoring 20% or more. Almost every argument about a highlighted sentence is really an argument about which of these two numbers somebody has in mind.

It is worth being precise about what this floor is and is not. An ai detector chosen by a writer is a private check; ai detection arriving through an institution is something else entirely, because a result can become a disciplinary matter without either party having chosen the tool. That difference is why this floor spends its time on how a number should be read rather than on how well the ai checker performs.

The second fact is even more important than the numbers, and it comes from Turnitin directly: the company states that it does not decide whether misconduct occurred. It supplies data; the judgement belongs to a person.

Where to start

Four ways in.

“I have been flagged and I wrote it myself.”
Start at if you have been flagged
“I am an instructor looking at a report.”
Read reading the report
“What do the highlighted sentences mean?”
That is reading the report
“What does Turnitin itself say about all this?”
Go to what turnitin publishes

What Turnitin publishes

Quoted from Turnitin’s own posts, with the dates. This wing exists because the company’s published position is considerably more careful than the way its number is often used.

It does not determine misconductTurnitin: “Turnitin does not make a determination of misconduct… rather, we provide data for educators to make an informed decision”.Being built
The document rateLess than 1%, and the condition attached: documents with 20% or more AI writing.Being built
The sentence rateAround 4%, higher in documents mixing human and AI writing, “particularly in the transitions”.Being built
Where the model goes wrongTurnitin: 54% of the time, a wrongly highlighted sentence sits right next to actual AI writing.Being built
What the company tells instructors“Use the information to initiate a conversation, not to draw a conclusion.”Being built

Reading the report

What the interface actually shows, and how to read it without importing certainty that the number does not carry.

The percentageWhat the figure is measuring, and the difference between a document score and a sentence highlight.Being built
Highlighted sentencesTurnitin asks that these be treated as “areas of interest”, not as findings.Being built
Low scoresWhy the smallest percentages are the least reliable, and what the company does about that.Being built
Similarity is a different reportThe older plagiarism score and the AI score measure unrelated things and are constantly confused.Being built

If you have been flagged

Written for the student, on the assumption that they wrote it themselves — which, given the published rates, is often true.

What evidence helpsDraft history, version records and notes — assembled before replying, not after.Being built
The conversationTurnitin’s own advice is that instructors should “assume positive intent” and give students “the strong benefit of the doubt”.Being built
Writing that gets flagged morePlain, structured prose, and the documented concern about writers working in a second language.Being built
Protecting yourself in advanceThe one habit — writing where the edit history is kept — that settles this before it starts.Being built

For institutions

Policy rather than technology. The published rates make some policies defensible and others indefensible, and this wing does that arithmetic.

A score is not a thresholdWhy a rule that acts automatically above a percentage — here or in GPTZero — contradicts what the vendors themselves say.Being built
What to tell students firstTurnitin’s advice to acknowledge false positives upfront, because not doing so makes the conversation worse.Being built
Assessment designThe changes that make detection largely unnecessary, which is where this argument ends.Being built
Keeping recordsWhat an institution should be able to show if a decision is challenged.Being built

What this tower will not do

It will not tell you a Turnitin score proves anything. The company says so itself, in plainer language than most of its critics use: it provides data, and the instructor applies judgement.

It will not rank Turnitin against other detectors. Nobody outside these companies can run that test properly, and the published benchmarks all come from the firms selling the product.

And it will not take a side against teachers. The instructor looking at a report did not build the tool, was rarely trained on it, and is being asked to make a judgement the vendor explicitly declines to make. That is a bad position to be put in, and several floors here exist for them rather than for the student.

One last note on sourcing, because it is unusual here. Everything quoted on this floor was read directly from Turnitin’s own blog rather than from a summary, and both posts carry the name of the company’s Chief Product Officer. Where a figure has a condition attached — and the headline one does — the condition is quoted with it, because dropping it is how a careful published statement turns into a number that decides somebody’s degree. What holds instead is simple: every figure and every quotation on this floor comes from Turnitin’s own published posts.

Where this page got its facts

  1. Turnitin on understanding false positives in its AI writing detection — the under-1% rate, and that it does not determine misconduct — www.turnitin.com, read 21 August 2026.
  2. Turnitin on the sentence-level false positive rate — around 4%, and the 54% proximity figure — www.turnitin.com, read 21 August 2026.

Written by Alberto Gulotta

Founder and editor of AI Tools Primer, writing from Palermo, Italy. Thirty-five years of taking computers apart, starting with a Commodore 64 — the long version is on the about page.

Something wrong on this page? Write to aitoolsprimer@gmail.com and it gets fixed.

Independence and limits

No affiliate links and no paid placements anywhere on this site. Nobody pays to appear here, and no company has seen this page before you did.

This is general information, not professional advice. Where a page touches money, health, safety or the law, it names its source and the date it was read — and your situation may still differ. See the privacy page and the cookie policy.