How to Read a Turnitin AI Writing Report: The Highlighting, Percentage, and Asterisk

You opened the report, saw a number in a colored box, some sentences glowing blue, and maybe a little asterisk you’ve never noticed before. Now you’re trying to figure out whether that means trouble. Here’s exactly what each piece of a Turnitin AI writing report is telling you — and, just as important, what it isn’t.

Key takeaways

• The big percentage is the share of qualifying text Turnitin flagged as likely AI-generated — not the odds you cheated, and not a plagiarism number.
• Blue highlighting marks the specific sentences attributed to AI; unhighlighted text was read as likely human.
• The **asterisk (\*)** shows up for scores between 1% and 19%. In that band Turnitin hides the exact number and the highlights, because false positives are more likely there.
• A blank score or “not enough text” message means your submission fell short of Turnitin’s 300-word minimum of qualifying prose.
• The AI score and the Similarity score are different systems and often disagree. Mixing them up is the most common error.

The three things people get wrong first

Before we go piece by piece, clear these up, because almost every panicked email to a professor starts with one of them.

The AI percentage is not the Similarity percentage. Similarity (the rainbow-colored matching report) compares your text against a huge database of published work and student papers. The AI indicator does something completely different — it estimates whether your writing *pattern* looks machine-made. You can score 0% similarity and 80% AI, or the reverse.

The percentage is not a probability that you’re guilty. A 60% does not mean “60% chance this student used AI.” It means roughly 60% of the analyzed text was flagged.

And the report is not evidence on its own. Turnitin says so directly: the indicator is meant to support a conversation, not replace human judgment.

The percentage: what the number counts

The number in the box is a proportion. Turnitin breaks your document into segments, scores each one, and the headline figure is the percentage of *qualifying* text it believes was AI-generated.

“Qualifying” matters. The tool only analyzes long-form prose. Bullet points, tables, code blocks, quotes, citations, and very short passages get excluded from the calculation. So if you submitted a 1,000-word essay but 300 words were block quotes and a reference list, the percentage is computed against the remaining ~700 words — not the whole file.

Take a concrete case. Maya submits a 1,400-word lab report and Turnitin shows 35%. That doesn’t mean 490 words are robotic. It means about 35% of the *prose Turnitin chose to analyze* landed above its internal AI threshold. Her methods section — dense with technical phrasing and predictable structure — is doing most of the flagging, while her discussion section, full of hedges and personal interpretation, reads as human. The single number hides that breakdown. The highlighting reveals it.

The highlighting: where the score lives

This is the part actually worth your attention. Turnitin highlights, in blue, the specific sentences it attributes to AI. Everything left plain is text the model considered likely human.

Read the highlights, not just the percentage. Two papers can both score 40% and tell totally different stories. In one, the flagging is spread thinly across every paragraph — often a sign of a consistent, slightly mechanical writing style, which non-native English writers and habitually formal writers both produce naturally. In the other, a single block is lit up blue and the rest is clean, which fits better with a chunk that got pasted in.

What the highlighting tends to catch: sentences with low variation, very even rhythm, generic transitions, and that smooth, agreeable, slightly hollow tone. If you want the statistical reason those patterns get flagged, this breakdown of why two detectors can score the same paragraph differently walks through perplexity and burstiness — the two measures most detectors lean on.

One honest limitation: Turnitin won’t tell you *why* a given sentence was flagged. There’s no per-word explanation, no confidence score on individual lines. You get the highlight and nothing more. That opacity is a real weakness, and it’s exactly why your own process evidence matters more than the colors.

The asterisk: the quiet caution flag

The asterisk is the part nobody explains, so here it is.

When the AI score falls between 1% and 19%, Turnitin doesn’t show you the exact figure or the highlights. It shows an asterisk instead — just *%, with no number attributed. This is deliberate. Turnitin’s own documentation says false positives are more likely at the low end, so rather than put a shaky small percentage in front of a reader who might over-read it, the tool withholds it.

So an asterisk isn’t a precise score you can act on. Read it as: “We picked up a faint signal that could resemble AI writing, but we won’t stand behind a number.” It’s the tool admitting uncertainty — the opposite of the overconfidence that gets students wrongly accused. If anything, an asterisk is closer to a non-result than to a flag.

“Not enough text” and the empty score

Sometimes there’s no percentage at all — just a dash, an asterisk in place of a number, or a message saying the submission couldn’t be processed.

Usually that means too little text. Turnitin’s AI checker needs at least 300 words of qualifying prose to produce a result. Below that, the signals are too sparse to score reliably, so it declines rather than guessing. Short reflections, discussion-board posts, and problem sets often come back blank for this reason.

Other triggers: a file that’s mostly quotes or references, text in a language the model doesn’t support (it currently handles English, Spanish, and Japanese), or formatting that strips out the analyzable prose. A blank score is not a pass and not a fail. It’s “we couldn’t say.”

Reading a real report start to finish

Picture Jordan’s report. Headline: 52%. Similarity: 4%. No asterisk.

Step one — separate the two numbers. The 4% similarity is fine; it’s a couple of properly cited phrases. Ignore it for AI purposes.

Step two — the 52% is well above the asterisk band, so it isn’t a suppressed low-confidence result. Worth taking seriously, not panicking over.

Step three — open the highlighting. Jordan sees the introduction and conclusion lit up solid blue, while the body paragraphs with specific examples and citations stay clean. That pattern says something: the framing language is generic and even, the substance is specific and varied.

What Jordan should *not* do is treat 52% as proof of anything. What Jordan *should* do is be ready to show drafts, notes, and version history — the actual record of how the work came together. When detection and process evidence conflict, process evidence is the stronger card.

Why none of these numbers are the last word

Worth remembering: detectors output probabilities, not proof. OpenAI retired its own AI-text classifier in July 2023, citing a low rate of accuracy — by its own report, the tool correctly identified only about a quarter of AI-written text. A 2023 Stanford study (Liang et al., published in *Patterns*) found several GPT detectors were biased against non-native English writers, misclassifying a large share of genuine TOEFL essays as AI-generated while rarely making the same mistake on native-speaker samples. Turnitin’s tool is more refined than those early classifiers, but it’s the same species of model, and it carries the same kind of risk.

That’s why the report is built as an *indicator*. The percentage, the highlighting, and the asterisk together are a prompt for a human to look closer — not a switch that decides anything on its own. If you write in a naturally even, formal style, you can light up a report without ever touching AI. Knowing how the report reads helps you make sense of a flag and respond calmly instead of defensively. And if you want to see what a checker reads in your own draft before you submit anywhere, you can look at the sentence-level breakdown first. There are also more guides on detection and writing if you want to go deeper on any one piece.

Frequently asked questions

What does the percentage on a Turnitin AI report actually measure? It’s the percentage of the qualifying text — long-form prose, with at least 300 words needed before the tool scores anything — that Turnitin attributed to AI generation. A 40% means about 40% of the analyzed writing was flagged, not a 40% chance you cheated and not a 40% source match. It’s an estimate from a probabilistic model, which is why Turnitin calls it an indicator.

What does the asterisk next to my AI score mean? It appears when the score lands between 1% and 19%. In that range Turnitin doesn’t show the exact number or the highlights — you just get an asterisk — because false positives are more likely there. Read it as a faint, low-confidence signal, not a verdict.

Why does my report say there isn’t enough text? The AI checker needs at least 300 words of long-form prose to score. Short answers, lists, code, and heavily formatted submissions get skipped. Fewer signals means the tool would rather say nothing than guess.

My AI score and Similarity score are totally different — which is right? Both, because they measure different things. Similarity matches your text against existing sources; the AI indicator estimates whether the writing pattern looks machine-made. They’re independent and are supposed to disagree sometimes.

Can I see why specific sentences were highlighted? Not in detail. The blue highlighting shows *which* sentences were flagged, but not the reasoning behind each one. That opacity is a genuine limitation — which is why your draft history and process notes carry more weight than the highlighting alone.

Bottom line

A Turnitin AI report has a few moving parts: a percentage that counts flagged prose, blue highlighting that shows where, an asterisk that stands in for a suppressed low-confidence score, and a blank result that means “too short to judge.” Read them together and the report stops being scary and starts being information.