Detexta

Turnitin AI Score Meaning: The Percentage, Explained

What the percentage is counting

In Turnitin's own description of its AI writing indicator, the headline number is about sentences, not words. It reflects the share of the document's sentences that the model has flagged as likely generated. That is a crucial detail, because it means the score does not scale with length the way people assume: a 20% score on a 2,000-word essay is roughly one in five sentences, which could be three paragraphs, not "one fifth of my work".

Institutions set their own thresholds for what counts as worth a conversation. Nothing is standardised, and a number that raises a question at one school is ignored at another.

The four things a score is made of

  1. The model's confidence per sentence. Each sentence gets a probability. The headline number is an aggregate of those.
  2. The threshold for flagging. A sentence above the cut-off is marked; below it is not. A 0% usually means nothing cleared the threshold, not nothing was detected.
  3. The version of the model. Detectors have been retrained repeatedly. A file scored in one term can score differently in the next.
  4. What was submitted. PDF versus DOCX, whether tables and references were included, and whether the text was in a supported language all change the result.

Why 20% is not one fifth of your essay

Two reasons. First, the unit is sentences, so a single long 60-word sentence counts the same as a short 8-word one. Second, flagged sentences cluster. A model-written paragraph usually produces a run of consecutive flagged sentences rather than a scattered fifth. So when you open the report, look for blocks — a block of red means one passage to fix, not a diffuse problem.

What to do with each band

Band What it usually means What to do first
0–10% Isolated flagged sentences, often formulaic transitions or standard definitions Read the flagged sentences; rewrite any that are doing real argumentative work
10–30% One or two flagged passages, or a document with several formulaic sections Find the longest consecutive run and rewrite it by hand, from notes
30–60% Large portions of the document share the model's texture Do not patch sentence by sentence; rewrite the structure of the flagged sections
60%+ The document reads as generated almost throughout Rewriting will take longer than writing again

These are descriptions of what the number looks like, not rules about what happened. A heavily formulaic lab report written entirely by a human can land in the middle bands.

The false positive problem, stated honestly

Style-based detection has a known weakness: writing that is genuinely formulaic looks generated. That includes technical reports, legal boilerplate, IELTS-style essays, and the work of writers composing in a second language. Several published studies have found elevated false positive rates in exactly these groups. Vendors acknowledge the limitation and position the score as one signal among several.

That does not mean the tool is useless. It means a score is a reason to look, and a human is supposed to do the looking. If you are on the receiving end of a flag, the useful response is documented drafts and a conversation, not an argument about the model.

Using a score on your own draft

Before submitting, the practical loop is:

A local reading of sentence variance, vocabulary spread and connector density will not match an official number, and it is not supposed to. It tells you where the texture is uniform, which is where you should be writing differently.

Frequently asked

Does 0% mean it found nothing? It means nothing cleared the threshold. Different tools apply different thresholds, which is why the same file can read 0% on one and 40% on another.

Can the score change without me editing anything? Yes, if the model version changes. That is one reason institutions treat it as a prompt for review rather than a result.

What score is safe? There is no universal number. Ask what threshold your institution applies. Most do not publish one, which tells you how much weight it is meant to carry.

Does the percentage include my references? Usually excluded, along with quotes, if your institution configured it that way. Check the settings on the report itself.