What Your Turnitin AI Score Actually Means

The report says 43% AI. Nobody tells you 43% of what. Not your professor, who often only sees the same number you do. Not the report itself, which offers a percentage and a blue highlight and very little else. So students end up guessing, and the guesses are usually worse than the truth.

I spend a lot of time walking students through these reports after the fact, so here is the plain version of what that number is, how it gets made, and what each range actually tends to trigger.

What the percentage measures (and what it doesn't)

The score is Turnitin's estimate of how much of your qualifying text came from an AI model. What counts as qualifying? Prose sentences in long-form writing, and only those. Your bullet points never get scored. Headings don't either, or tables, or code, or the citations at the end. So a 43% doesn't mean 43% of your document. It means the model looked at your eligible sentences and decided 43% of them read as machine-made.

Just as important, it is not a confidence level. A 43% does not mean Turnitin is 43% sure you used AI. The model scored each eligible sentence, decided a chunk of them looked machine-generated, and that chunk added up to 43% of the qualifying prose. Those are very different claims, and the first one gets students disciplined for the second one all the time.

How the number gets made

Mechanically, here's what happens after you hit submit, per Turnitin's own documentation:

• Your paper needs at least 300 words of prose (and under 30,000) in English, Spanish, or Japanese, or no AI report generates at all.
• The text gets chopped into overlapping segments of roughly five to ten sentences, so every sentence is evaluated in context.
• Each sentence gets a score between 0 and 1: how strongly it resembles language-model output.
• Those sentence scores aggregate into the single percentage you see.

Notice what's missing: nothing in that pipeline knows whether you used ChatGPT. There's no log, no metadata, no smoking gun. It's a statistical judgment about how predictable your sentences look, sentence by sentence.

Why low scores show an asterisk instead of a number

If your score lands anywhere from 1% to 19%, Turnitin doesn't display the number. You get an asterisk, because the company's own testing found that scores in that range are wrong too often to be worth showing. That's not a criticism from me, it's their published position: below 20%, the false positive risk outweighs the value of the number.

Sit with that for a second, because it's the most useful fact in this whole report. The vendor itself draws a bright line and says: under this threshold, don't trust me. The threshold where they do trust the number is a judgment call, and their acknowledged sentence-level false positive rate of roughly 4% still applies above it.

What each range tends to trigger in practice

• 0%: the model found nothing. This is most human writing, most of the time.
• Asterisk (1–19%): officially "unreliable." Most instructors are trained to ignore it. A few unfortunately don't.
• 20–40%: the gray zone. This is where formal-but-human writing, heavy grammar-tool use, and light AI assistance all blur together. Usually prompts a conversation, not a charge.
• 40–80%: serious scrutiny. Expect to be asked about your process, and expect your draft history to matter.
• 80–100%: the model believes most of the paper is machine-generated. At this level, cases tend to move to formal review quickly.

One caveat that surprises people: these ranges describe what typically happens, not what your school's policy says. Turnitin's own guidance is explicit that the score should never be the sole basis for an integrity decision. Universities that trust that guidance treat every score as the start of a conversation. Universities that don't, don't.

And some schools have opted out of the argument entirely. Vanderbilt, MIT, and Yale all turned the AI detector off campus-wide rather than adjudicate its gray zones. If your school still runs it, your score exists inside a policy, so find that policy before your meeting, not during it.

What your instructor sees that you don't

Here's the part that makes these conversations so lopsided: in most course setups, students never see the AI report at all. The percentage, and the sentence-level view behind it, are instructor-facing. You find out your own score secondhand, in an email or a meeting, usually without the context that comes with it.

What the instructor sees is your paper with the flagged passages highlighted. Two colors, in the current report. One marks text the model thinks came straight out of an AI. The other marks text it thinks was AI-generated first and paraphrased after. That second color matters. It exists because students kept running ChatGPT output through rewording tools, and it means the but-I-paraphrased-it defense doesn't work the way a lot of students assume.

This is also why asking your instructor which sentences were flagged is such a reasonable request. The information exists, it's sitting in their view of the report, and a percentage without it is almost impossible to respond to meaningfully. Most instructors will show you when asked directly.

The AI score is not the similarity score

Turnitin produces two unrelated numbers, and students mix them up constantly, sometimes mid-defense. The similarity score is the old one. It measures how much of your text matches existing sources, and quoting or citing correctly still produces a nonzero number there. That's normal. The AI writing score is the newer one, and it's what this article is about. A paper can be 2% similarity and 90% AI. It can be 40% similarity and 0% AI. If you're going to argue about one of them, first make sure you know which one you're arguing about.

If the number is high and you wrote the paper

Formal academic prose is exactly the register these models flag most, which is why polished human writing lands in the gray zone so often. Grammar tools push in the same direction, and if that's part of your situation, the Grammarly-versus-Turnitin question has its own specific answer depending on which features you used.

If an accusation is already on the table, the score itself is not something you can argue down, but it's also not something they can convict on alone. What wins is process evidence: version history, dated drafts, notes. There's a step-by-step playbook for proving you wrote it that covers exactly what to gather and in what order.

And if you're reading this before the deadline rather than after: run the paper through a free AI detector first. It won't predict Turnitin's exact number, no external tool can, but it will tell you whether your writing sits in the range that invites questions, while you still have time to revise it.

The number is a screenshot, not a verdict

Turnitin's AI score is one model's opinion about the statistical texture of your sentences, measured only on the parts of your paper it considers prose, hidden entirely when it's below the vendor's own reliability line. That's not nothing. It's also nowhere near proof, and the company that built it says so in writing.

Know which number you're looking at, know what it was measured against, and keep your drafts. Students who understand the report almost always come out of these conversations better than students who just fear it.