What Percentage of AI Detection Is Allowed in Turnitin?

Turnitin does not set an allowed percentage, and neither does any national body. Institutions set their own rules. In practice Turnitin suppresses every score between 1 and 19 percent, showing an asterisk instead of a number, and most universities with a written policy treat anything under 20 percent as unremarkable while starting a manual review somewhere in the 20 to 50 percent band. The number itself is never the decision. It is a prompt for a human to look closer.

That answer frustrates people, because everyone wants a line. The reason there is no line is that Turnitin publishes the score as a probability estimate rather than a finding, and it says so plainly in its own documentation. What follows is what the bands actually mean, what different institutions do with them, and which things quietly inflate a score that has nothing to do with how you wrote.

What percentage of AI detection is allowed in Turnitin?

There is no universal allowed percentage. Turnitin reports a score and leaves interpretation to the institution, which is why the same 28 percent can be ignored at one university and trigger a meeting at another. The one rule Turnitin does enforce itself is the suppression band: readings from 1 to 19 percent are not attributed at all, because the company considers them too unreliable to publish.

Read that table as a description of habit, not law. Some departments have no threshold written anywhere and rely on whoever is marking. Some have switched the AI indicator off completely because of false positives. A handful publish an explicit tolerance in the module handbook. If yours has one, it is the only number that matters to you, and it is worth finding before you need it.

Why does Turnitin hide scores between 1 and 19 percent?

Turnitin suppresses the 1 to 19 percent band to avoid false positives, because at those levels the model cannot distinguish AI-like human writing from actual AI writing with enough confidence to report a figure. Short passages of formal, tightly structured prose look statistically similar to generated text, and plenty of good academic writing is exactly that.

This is the single most useful thing to understand about the score. The vendor that built the detector decided its own low readings were not trustworthy enough to show you. That tells you how to read every other detector's low numbers too, and it is why a free checker returning "18 percent AI" means close to nothing.

What percentage of AI is acceptable in college?

Most US colleges that have published a policy accept AI assistance for research, outlining, grammar and feedback, and prohibit submitting generated prose as your own work, regardless of what percentage a detector reports. The policies are written about conduct, not about scores. A 30 percent reading on work you wrote yourself is not misconduct, and a 4 percent reading on an essay ChatGPT wrote is not innocence.

In practice you will meet three regimes. Some instructors ban AI outright for a given assignment. Some allow it with disclosure, asking you to note where you used it. Some do not mention it at all, which usually means the institutional default applies. Departments that handle this well write the threshold down, put it in the course handbook, and run markers through the same short briefing on how to read an AI report, which is exactly the sort of guidance an internal training and certification platform exists to standardize. Ask before you submit, not after.

Is 20 percent AI detection bad?

A 20 percent AI score is the lowest number Turnitin will actually show you, so it reads as worse than it is. It means roughly a fifth of the qualifying text carried AI-like statistical patterns, which happens routinely in methods sections, literature reviews and any passage written in a flat, formal register. On its own it is not evidence of anything.

What makes 20 percent uncomfortable is the visibility jump. At 19 percent you get an asterisk and nobody looks. At 20 percent a number appears next to your name along with highlighted sentences, and highlighted sentences invite reading. If your writing sits naturally in that zone, and formal academic prose often does, the sensible move is to know which parts get flagged and why, rather than to wait and hope.

Does Turnitin flag citations and references as AI?

Turnitin's AI model evaluates long-form prose and is not supposed to score reference lists, block quotations or citation strings, but the AI percentage is calculated against qualifying text only, which can push the reported figure up on citation-heavy work. A paper that is 40 percent quotations and bibliography has a much smaller pool of prose being judged, so a few flagged paragraphs occupy a larger share of it.

This catches people out on literature reviews in particular. The prose between citations tends to be formulaic by convention, since you are summarizing other people's findings in a neutral register, and neutral register is the most AI-like thing a human writes. It is not a bug you can argue away, but it is worth knowing when your score comes back higher than the effort you put in suggests it should.

Does Turnitin flag Grammarly as AI?

Turnitin scores the text you submit, not the tools that touched it, so Grammarly's spelling and grammar corrections do not register. Grammarly's generative features are different: accepting a rewritten sentence or a generated paragraph inserts machine-written prose, and that is what a detector reads. The line falls between fixing what you wrote and having something write for you.

The other half of this is Grammarly Authorship, which records how text arrived in a document rather than analyzing the finished words. That is provenance, not detection, and it behaves completely differently from a statistical score. We covered the distinction in detail in how Grammarly's AI detection actually works.

How accurate is the score you are being judged on?

Turnitin claims 98 percent accuracy with a false-positive rate under 1 percent on documents scoring above 20 percent AI. Independent testing puts real-world accuracy closer to 90 to 95 percent on untouched model output, with false positives between 5 and 12 percent on non-native English, technical prose and heavily edited drafts. A Stanford-led study found that widely used detectors misclassified non-native English writing as AI at rates above 60 percent.

Those numbers are the reason several universities, including Vanderbilt, turned the feature off rather than defend it case by case. They are also why the score is a starting point for a conversation and not the end of one. If you want the mechanics of what the model measures and where it breaks, our breakdown of how Turnitin detects AI goes into the detail, and the wider question of detector accuracy covers the rest of the field.

What to do if your score comes back higher than you expected

First, do not panic and do not rewrite in a hurry. A score is not an allegation, and rushing produces worse writing that still scores badly. Work through this instead.

• Find your institution's actual policy. Module handbook, then the academic integrity page. If a tolerance is published, you are arguing about a written number rather than a feeling.
• Look at which passages are highlighted. If they are your methods section and your literature summary, you are looking at register, not authorship, and that is a defensible position.
• Gather your process evidence. Document version history, notes, outlines, reading annotations and search history all show work happening over time. This is far more persuasive than any score you could produce.
• Vary the prose that flagged. Uniform sentence length and predictable transitions drive the reading up. Breaking that pattern is good writing advice independent of any detector, and it is what a humanizer built for Turnitin submissions does mechanically.
• Ask before you submit, not after. A short email to your instructor asking how they read AI reports costs nothing and changes the conversation entirely if a flag ever appears.

If a false flag has already landed on you, the practical steps for clearing a false Turnitin flag are worth reading in full, because the appeal succeeds or fails on evidence of process rather than on argument about the percentage.

Worth knowing before you go looking: you cannot check the number yourself beforehand. Turnitin restricts the AI indicator to instructors and administrators, so there is no public Turnitin AI checker and no way to compare a third-party score against the one your institution will see.

Is the percentage you see the model's real estimate?

No, and Turnitin says so. The reported figure is deliberately conservative, because holding the false positive rate under 1 percent costs the model recall. Its FAQ gives a worked example: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing." So the number on the report is a floor rather than a best guess.

This cuts both ways, and neither side usually likes it. If you are being reviewed, a 30 percent reading means the model's own estimate of the underlying figure is higher than 30. If you are the one doing the reviewing, a low number is weak evidence of anything, because the system is tuned to stay quiet when it is unsure. Turnitin's own instruction is that the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors".

One more thing changed recently enough that most advice on this topic is out of date. In July 2026 Turnitin consolidated what had been a multi-model ensemble into a single model, and states the change maintains the same sub-1 percent false positive rate. No independent test of that version exists yet. We traced every accuracy figure in circulation, including which ones have no published method behind them, in our breakdown of how accurate the Turnitin AI detector actually is.

The short version

No allowed percentage exists. Turnitin will not show you anything under 20 percent because it does not trust its own readings down there. Above 20 percent you get a number and highlights, and what happens next depends entirely on your institution and on what your document history says about how the work was made. Treat the score as weather rather than judgment: worth checking before you go out, not worth arguing with.

If you write in a formal register and keep getting readings you cannot explain, comparing the tools built for this problem is the practical next step. Our comparison of AI humanizers for Turnitin lays out what six of them cost and where each one loses, with prices we read off the vendors' own pages.