Two different checks, one confusing report
An AI checker and a similarity checker answer different questions, and almost every argument about a "Turnitin score" comes from mixing them up. The similarity side asks: does this text overlap with something already in a database? The AI side asks: does this text read like it was produced by a language model? You can score 0% similarity and 90% AI, or 30% similarity and 2% AI. They move independently.
This page is about the second one, and about what you can actually do with a draft before the deadline.
What an AI-writing model is looking for
Turnitin has published the shape of its approach rather than the model itself. The public description talks about the statistical regularity of generated text — the idea that a language model produces prose that is unusually even. Independent researchers have found the same signals again and again:
- Sentence lengths that barely vary. Human writing swings from four words to forty. Generated text tends to hover around a comfortable middle.
- Low burstiness. This is the technical name for that variance. Short clusters of punchy sentences followed by a long, winding one is a human fingerprint.
- Narrow vocabulary spread. Not simple words — repeated reliance on the same abstract nouns. Framework, landscape, crucial, robust, comprehensive.
- Transition phrase density. Moreover, furthermore, in conclusion, it is important to note. These are the connective tissue of a model that was trained to sound organised.
- Perfectly balanced hedging. Every claim softened by exactly the right amount, so nothing is ever wrong and nothing is ever said.
You can measure the first four of those in a browser tab, which is exactly what the reader at the top of this page does. It cannot measure the fifth, and neither can you, consciously.
Why the same essay passes one tool and fails another
Every vendor trains a different model on a different corpus and picks a different threshold for calling something AI. Add the fact that most vendors do not publish their threshold, and you get the situation students describe constantly: 4% on one tool, 67% on another, 0% on a third, for the same file.
Three things drive the disagreement:
- The reference corpus. A model trained heavily on academic essays will read a lab report differently than one trained on web text.
- The threshold. Some tools report the raw probability. Others apply a cut-off and only flag what clears it. A "0%" can mean nothing cleared the bar, not nothing was detected.
- Pre-processing. Whether bullet points, quotes and references are stripped before scoring changes the result a lot. A long block quote from a source can drag the surrounding style score around.
This is why the useful question is not "what is my score" but "which sentences are carrying the score". A single long, smooth, transition-heavy paragraph can hold up an entire document.
The order to fix things in
If your own check turns up red bars, work through them in this order. It is also the order that improves the writing for a human reader, which is the only reason it is worth doing.
- Split the longest sentence in each flagged passage. Not into two equal halves — into one short claim and one sentence of detail.
- Replace the transition with the relationship. Instead of moreover, write and the same thing happens when. Instead of in conclusion, say what you actually concluded.
- Name one concrete thing per paragraph. A date, a number, a person, a title. Generated text is full of categories; human text is full of instances.
- Keep one deliberate fragment or aside. A real writer breaks rhythm on purpose. Do not manufacture this — if you have a natural aside, leave it in.
- Re-run the check. Watch the bars move. If a sentence stays red after two rewrites, the problem is probably that the sentence is doing too much, not that it sounds like a machine.
What this cannot tell you
It cannot tell you what your institution's tool will report. It has no access to Turnitin's model, its database, or your submission history. It also cannot see the things a human instructor sees: whether you can explain the argument out loud, whether the citations match the sources, whether the voice is consistent with the rest of your work.
Treat the number as a hint about where to look, and the sentence list as the actual output.
Frequently asked
Is the AI score the same as the similarity score? No. Similarity is about matching sources. The AI score is about style. They are reported separately and they can point in opposite directions.
Does a high AI score prove someone used AI? No, and vendors say so themselves. False positives are common in technical, formulaic and non-native writing. The score is a signal for a human to investigate, not a finding.
Can rewriting get a genuine essay to zero? Sometimes, and that is a real problem with style-based detection. If you are rewriting your own work to satisfy a tool, keep the earlier drafts — they are the best evidence you have if it comes to a conversation.
Is my text uploaded when I use the reader on this page? No. It runs as JavaScript in your browser tab. Nothing is sent anywhere and nothing is stored.