Turnitin AI Report: What It Really Says and How to Reply

A Turnitin AI report is a probability estimate, not a verdict. It says a pattern in your text resembles model output. It does not establish that you used AI, when you used it, or what you used it for. What you do next matters more than the number on the screen.

Key takeaways

• The report gives you a percentage and often a sentence-level view; neither one is proof of authorship.
• A high score can come from your own writing habits, not from a chatbot.
• Before you reply to anyone, export the report, note the flagged sentences, and gather your drafts.
• Re-check the flagged sentences yourself against your own outline and notes.
• Fix the draft only if the flagged sentences genuinely read flat or generic.
• Keep the reply factual and short; do not argue about accuracy you cannot verify.

What changed in how Turnitin reports AI?

Turnitin added AI-writing detection to its similarity checking, and the way it presents results has shifted more than once. Early versions showed a single document-level percentage. Later versions moved toward sentence-level highlighting, and Turnitin has publicly acknowledged that the signal is weaker on shorter passages and on writing that has been edited.

That matters because most people read the first number they see and stop. If the report says a document is flagged, they assume the whole thing was generated. A sentence-level view tells a different story: a handful of lines carry the signal, and the rest of the document may look ordinary.

We ran our own detector on a set of hand-written DeAIze guides on 2026-09-19 (n = 46, local ensemble, lite tier, 30% flag threshold). The mean AI score was 27% and the median was 25%, with a range from 0 to 55%. Seventeen of those 46 human-written documents were flagged at that threshold. Human writing gets flagged. That is the part of the story the report itself does not tell you.

The same run scored 12 unedited model passages: mean 35%, median 34%, range 25-49%. The distributions overlap. A single score cannot separate the two groups cleanly, which is exactly why the sentence-level view and the context around your writing matter.

What does the Turnitin AI report actually establish?

It establishes that a statistical model found patterns it associates with generated text. That is all.

It does not establish intent. It does not establish that you pasted anything. It does not establish which tool was used, or whether a tool was used at all. It cannot see your browser history, your notes, or your drafts. It is a classifier reading text, and classifiers produce probabilities.

There are three common reasons a genuine draft gets flagged:

1. Formulaic structure. If every paragraph opens with a topic sentence and closes with a summary, the rhythm looks machine-made, because that is how models are trained to write.
2. Low burstiness. Human prose mixes long and short sentences unevenly. Uniform sentence length reads as generated.
3. A register mismatch. Non-native English writers who learned formal academic phrasing often produce text that detectors associate with model output. This is a known and documented pattern, not a personal failing.

If any of those describe your draft, the report is telling you something about style, not about honesty.

If you want to see which specific sentences carry the signal, run your draft through a sentence-level detector before you reply to anyone. DeAIze’s AI detector scores the whole document and highlights each sentence by AI probability, so your reply can point at lines instead of arguing about a number.

What should you do first, before replying?

Do not reply while you are still upset. Do these four things in order.

1. Save the report. Download the PDF or screenshot the flagged view. Note the date, the score, and which sentences are highlighted. If the score changes later, you want the original.
2. List the flagged sentences. Write them out in a separate document, numbered. This is the only part of the report that gives you anything to work with.
3. Pull your evidence. Drafts, outline, notes, reading list, earlier versions, timestamps. Anything that shows the work happened over time.
4. Re-read the flagged sentences cold. Cover the rest of the paragraph. Read each one aloud. Ask: does this sentence say something specific, or does it say something generic in a confident voice?

That last question is the useful one. Generic confidence is what detectors react to. If a flagged sentence could appear in almost any essay on the topic, it is doing no work, and that is a writing problem you can fix regardless of what the report says.

If you are accused formally, the sequence is different, and we cover the reply itself in our guide to false positives and what to do.

How do you re-check the sentences yourself?

You do not need a detector to do this. You need five minutes and a willingness to be honest about your own prose.

Take the flagged sentences and run them through four tests:

Test	What you are looking for	What a pass looks like
Specificity	A name, number, date, or concrete detail	“The 2019 field trial in Kent” rather than “recent research”
Claim	An assertion you could defend in a viva	A position, not a summary of a position
Rhythm	Variation in sentence length	A short sentence somewhere nearby
Voice	A choice only you would make	An aside, a hedge, a judgement

A sentence that fails all four is a sentence a detector will like and a reader will skim. That is worth fixing whether or not you were ever flagged.

Here is a worked illustration. The before version is the kind of line that shows up highlighted:

Before: “It is important to consider the various factors that contribute to this phenomenon.”

After: “Three things pushed the cost up: fuel, labour, and a shortage of drivers.”

The second version names the things. It is shorter. It commits to a claim. Nothing about the meaning changed; the vagueness did.

If you want a second opinion from a different model family, cross-checking with another tool is reasonable, but read our notes on why two detectors disagree first, because a second score is not a tiebreaker. It is just another estimate.

How do you fix the draft without losing your argument?

Rewrite only the flagged sentences. Not the whole document. This is the single most common mistake people make, and it is how good arguments get flattened.

Work in this order:

1. Keep your structure. The outline is yours. Leave it alone.
2. Rewrite flagged sentences only. Replace vague nouns with specific ones. Split a long sentence. Add one concrete detail you actually know.
3. Read the paragraph aloud. If the rewritten sentence no longer connects to the one before it, you changed too much.
4. Re-score. Run the revised draft back through the detector and compare before and after.
5. Check the meaning survived. If a rewrite changed your conclusion, revert it.

DeAIze’s humanizer works on this principle: it scores the text, rewrites only the flagged sentences, keeps a rewrite only when semantic similarity says the meaning survived, then re-scores and reports before and after. Across our own runs the site reports an average AI-score reduction of 60%+ after humanizing, best case 71%. Those are our own predicted scores, not a verdict from Turnitin or any other third-party detector, and no tool can promise you a particular score on someone else’s system.

On the example paragraph used on the site, a 92% AI score becomes 4% AI. That is one paragraph, under our own measurement, not a guarantee.

One honest caveat: if large parts of your draft were generated and you did not write the argument, rewriting sentences will not fix that, and it should not. The ideas, data and conclusions have to be yours. Humanizing is for reducing false positives on your own writing and for cleaning up your own AI-assisted drafts.

What should your reply actually say?

Keep it short, factual, and free of accusations in either direction. A workable shape:

• State that you wrote the work and can account for its development.
• Attach the evidence you gathered: drafts, notes, timestamps.
• Point at the specific flagged sentences and explain, briefly, why they read the way they do. If a sentence is generic, say so and offer the revision.
• Ask what process follows, and what standard is being applied.
• Do not claim the detector is broken. Do not cite accuracy figures you cannot verify. Detector output is a probability estimate, and saying so plainly is stronger than overclaiming.

If the conversation is about a specific tool, our breakdown of whether Turnitin detects ChatGPT covers the mechanics without the panic.

What the report cannot do, and why that matters

It cannot read your process. It cannot see the two hours you spent restructuring a paragraph. It cannot tell the difference between a student who pasted an answer and a student who writes in a clean, uniform style because that is what they were taught.

Our own measurement makes the limit concrete. In that 2026-09-19 run, hand-written documents averaged 27% and unedited model output averaged 35%. The gap is real but narrow, and the ranges overlap heavily. Any system that turns a number like that into a verdict is overreading its own instrument.

That is not a reason to ignore the report. It is a reason to treat it as one input among several: your drafts, your notes, your ability to explain your own argument, and the judgement of a human reader who actually looks at the work.

Ready to see which sentences read machine-made?

Before you reply to anyone, get the sentence-level view. Paste your draft into the AI detector and see which lines carry the signal, then open the draft editor to rewrite only those lines and re-score the result. If you want to compare plans first, pricing is pay-as-you-go, credits never expire, and there is no subscription. You can start with 200 words free on signup, no credit card. Upload a .docx and the rewrite preserves your fonts, styles and tables, so you are not rebuilding your formatting afterwards. And if you only take one thing from this page: the report is an estimate, your evidence is the argument, and the sentences are the part you can actually control.

---