Worried Turnitin will flag your writing? Or did it already?

How Turnitin's AI detector works

Turnitin has published more methodology than most competitors. What they describe is a transformer-based deep-learning classifier that processes documents in overlapping windows.

Segment windows

The classifier reads the document in segment windows of roughly a few hundred words (about 5 to 10 sentences). Windows stride across the document one sentence at a time, so each sentence ends up evaluated in the context of several windows.

Per-window probability

Each window produces a real number between 0 and 1, where 0 means the text is unlikely to be AI-generated and 1 means it is strongly plausible AI. The document score aggregates the window-level predictions into a percentage of qualifying text identified as AI.

Statistical and linguistic signals

Turnitin describes signals that include vocabulary distribution, sentence structure, and how information unfolds across paragraphs. These overlap conceptually with perplexity and burstiness, though Turnitin doesn't name them with those terms.

Model coverage

Turnitin's published coverage extends to GPT-3, GPT-3.5, GPT-4, GPT-4o, GPT-5, Claude Sonnet 4.5, Gemini, LLaMA, and tools built on these models. As of July 2024, the detector also identifies AI text that has been run through paraphrasing or word-spinning tools, highlighted separately in purple in the report. Paraphrasing detection is currently English-only.

Where Turnitin is more constrained: only English includes the AI paraphrasing and AI bypasser detection features. Spanish and Japanese submissions get the basic detector with reduced coverage. The system requires a minimum of 300 words and supports up to 30,000 words of qualifying text per report. Accepted file types are .docx, .pdf, .txt, and .rtf.

What your Turnitin AI score actually means

Unlike most consumer detectors, Turnitin publishes specific interpretation guidance and explicit display rules. The score is a percentage representing the portion of qualifying text the classifier identified as AI-generated.

The most important thing to know about Turnitin's display convention.

0% — no AI detected

The document is not flagged.

1–19% — asterisk (*) shown, no number

Turnitin intentionally suppresses the percentage in this range. Their published reasoning is that scores between 1% and 19% are less reliable, so showing the asterisk avoids implying false precision to instructors. The asterisk is itself a signal that the result should be treated cautiously.

20–100% — percentage shown

At 20% and above, the actual percentage is displayed. Turnitin reports a document-level false-positive rate of less than 1% in this range, validated against 800,000 pre-GPT documents.

“We have discovered real-world use is yielding different results from our lab.”

That admission, made about two months after the detector launched, led to the asterisk convention and the increase in the minimum word count from 150 to 300. It is unusually direct for a commercial detector and worth taking seriously: even Turnitin treats scores in the lower range as uncertain.

Known failure modes of Turnitin's detector

Turnitin has been more transparent about its limits than most commercial detectors. The patterns below come from their own published statements and from independent research.

Scores in the 1–19% range

Turnitin's own reasoning for the asterisk is that this range has higher false-positive rates. If your instructor sees an asterisk, that is by design a signal of uncertainty, not a quiet accusation.

Document beginnings and endings

Turnitin acknowledged in 2023 that false positives were concentrated at the beginning and end of documents and adjusted their sentence aggregation accordingly. Edge effects are still worth being mindful of when reviewing flagged sentences.

Short submissions

Documents under the 300-word minimum may not be evaluated at all, and even just-above-threshold documents can produce less reliable results. Longer text gives the segment-window classifier more context.

Formal academic register

Third person, passive constructions, hedged claims, citations. The patterns that academic writing rewards (low surprise, consistent vocabulary, structured argument) are the same patterns the detector reads as AI-like. Conscientious writers are systematically more likely to land in the asterisk range than casual writers.

Non-native English patterns

Writers whose first language is not English often rely on more structured phrasing and a more limited vocabulary, which can lower perplexity-like signals. This is a documented bias across every commercial AI detector, not specific to Turnitin.

Spanish and Japanese submissions

The basic detector works on these languages, but the AI paraphrasing and AI bypasser detection capabilities are currently English-only. If you submit in a non-English language, the report you get is less complete than an English submitter would.

What to do if Turnitin flagged you

Because Turnitin is the institutional detector, “flagged” here usually means an instructor has already seen a score and opened a conversation with you. The steps below assume that context.

01Read the report carefully

Look at what is highlighted, not just the percentage. Turnitin shows the specific sentences flagged. AI-generated content shows up in one color and AI-paraphrased content (post-July 2024) shows up in purple. The pattern of highlights matters more than the headline number.

02Note whether the score is asterisked

An asterisk means Turnitin itself has acknowledged the result is less reliable. That is your most important rhetorical lever in an honest conversation with your instructor.

03Document your writing process

Save drafts, version history, research notes, outline drafts, and timestamps. Google Docs and Word both keep version history automatically. If a real accusation develops, the timeline of your work is what protects you.

04Ask for the specific sentences in question

In a conversation with an instructor, the specific flagged passages are more discussable than the document-level score. Asking “which sentences are you concerned about?” shifts the conversation from a number to your actual writing.

05Rewrite with intention next time

If your writing keeps landing in the asterisk range, the issue is usually structural: low sentence-length variation, formal register, predictable vocabulary. Vary your sentence rhythm. Drop in a hedge, an aside, or a contraction. Plain prose almost always scores lower than the same content in tight academic voice.

For the longer post-flag playbook, see “Your work was flagged as AI. Here's what to do.”

What GPTypo gives you that Turnitin doesn't

Turnitin's detector is institutional. You cannot scan your own work through it; you only see the score after your instructor does. GPTypo runs the same kind of analysis at the sentence level and lets you preview before submission. We don't try to fool the detector. We surface the risk so you can edit in your own voice before it becomes a problem.

✓Pre-submission previewing

Run your draft and see which sentences would draw flags before your instructor opens the Similarity Report. The scoring is independent and fully under your control.

✓Sentence-level scoring

Every sentence is scored independently. You see exactly which 20% of your writing is driving 80% of the risk, instead of staring at a single document-level number.

✓Rewrite suggestions on demand

Click a flagged sentence to get two rewrite options that preserve your meaning and reduce the signals the detector reacts to.

✓Independently benchmarked

Our scoring model is on the public RAID leaderboard at raid-bench.xyz; methodology at /verification.

AI detection with sentence-level rewrites. Independently benchmarked.

How Turnitin compares to other AI detectors

Reading the table. Turnitin is the only detector on this list with institutional access at the level it has, which is what makes its results consequential. None of these detectors are reliable enough to use as the sole basis for an accusation. All have meaningful false positive rates on writing that overlaps with AI patterns (formal register, ESL, structured documents, short text).

Common questions about Turnitin's AI detector

Last updated: May 3, 2026. Turnitin actively updates its detector capabilities; the most recent material change at the time of writing was the November 2025 admin CSV-export feature. The methodology, score display rules, and false-positive figures cited above reflect Turnitin's published statements through 2025.