Updated August 2026 · read from Turnitin's own docs
How accurate is Turnitin AI detector? Turnitin AI checker accuracy and false positives
Turnitin publishes a false positive rate of under 1%. It attaches a condition to that number which almost nobody quotes, and it hides every reading between 1% and 19% from the person being graded.
Below: what Turnitin's own documentation says, what the small amount of outside testing found, and which widely repeated accuracy figures have no traceable source.
2 free runs a day, up to 200 words each. We save your run so you can get back to it, and a delete button appears with the result. See privacy.
This is our own AI-pattern score, measured here on sentence rhythm, template phrases, vocabulary variety and passive voice. It is not a GPTZero, Turnitin, Originality.ai, Copyleaks or ZeroGPT result, and it does not predict one. Worth knowing: we also ask the rewrite to vary sentence length, drop template phrases and prefer the active voice, so some of the drop is built in. Read the two panels below, not just the number.
Before ·
After ·
Your own text rewritten · Saved to your history, delete any time
The short answer
Turnitin's AI detector reports a false positive rate of under 1%, but only for documents it has already scored above 20% AI, and it validates that figure internally against more than 700,000 academic papers written before ChatGPT existed. Independent testing barely exists. The two outside checks on record, a Washington Post test in April 2023 and an Inside Higher Ed report in February 2024, both measured a model two architecture changes out of date. So the honest position is that Turnitin is probably the most carefully calibrated AI detector in education and that nobody outside the company has verified it, including for the single model that matters, the consolidated one it shipped in July 2026.
Last updated August 12, 2026. Every Turnitin figure below was read from turnitin.com and guides.turnitin.com on that date. Where a number could not be traced to a method, sample and date, this page names it and declines to repeat it.
Claim versus source
Every Turnitin accuracy figure, and where it actually comes from
Six numbers circulate about this detector. Four come from Turnitin, two come from outside it, and none of them means what a headline suggests on its own.
| The figure | Who published it | What it leaves out | Status |
|---|---|---|---|
| Under 1% false positives | Turnitin's own AI writing detection FAQ | Applies only "for documents with over 20% of AI writing". The rate is not published for documents below that line, which is where a wrongly accused student actually sits. | Vendor claim, conditional |
| 0.51% at document level | Turnitin's own whitepaper | A second, lower number from the same company. Both are internally produced. No outside party has audited either. | Vendor claim, unaudited |
| Validated on 700,000 papers | Turnitin's own FAQ | Verbatim: before every model release it tests "over 700,000 additional academic papers that were written before the release of ChatGPT". A genuinely large pre-ChatGPT control set, and more disclosure than any competitor offers. | Method disclosed |
| About 50% accuracy | Washington Post, April 1 2023 | A small-sample press test on an early version of the model, run 16 months before the July 2026 architecture change. Old, and too small to generalize from, but it is one of the very few outside tests that exists. | Independent, small sample |
| Missed roughly 15% of AI text | Inside Higher Ed, February 2024 | A false negative finding rather than a false positive one. Consistent with Turnitin's own stated trade-off, which is to accept misses in order to hold the false positive rate down. | Independent, dated |
| 61.3% misclassification | Liang et al., Patterns, 2023 | Seven detectors on TOEFL essays by non-native English writers. Turnitin was not one of the seven, so this does not measure Turnitin. It measures the category-wide failure mode that Turnitin says it trained against. | Peer reviewed, not Turnitin |
The conditions on the 1%
Six things Turnitin's own documentation says about its accuracy
None of these is a leak or an accusation. All six are published by Turnitin, and all six change what the headline rate means.
The 1% only covers documents already over 20% AI
Turnitin states its target is keeping false positives "under 1% for documents with over 20% of AI writing". Read that condition carefully. The guarantee is scoped to documents the model has already decided are substantially AI. For a paper you wrote yourself, which is the case that matters, the published rate simply does not apply.
Anything between 1% and 19% is hidden from everyone
Turnitin suppresses its own reading in that band. Verbatim: "no score or highlights are attributed for AI detection scores in the 1% to 19% range." An asterisk appears instead of a number. The stated reason is its own testing, which "found that there is a higher incidence of false positives when the percentage is between 0 and 19".
The score deliberately under-reports
Holding false positives down costs recall, and Turnitin publishes the size of that cost: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing." The number an instructor sees is a floor, not an estimate.
A 1% document rate is not a 1% student rate
One in a hundred sounds small until it is multiplied by submissions. A course of 300 students submitting four papers each produces 1,200 documents, and at one percent that is roughly a dozen false flags a year in a single course. This is arithmetic, not a criticism of the model, and it is why no vendor rate should be read as "this will not happen to me".
It is not a verdict, by Turnitin's own instruction
Turnitin writes that it "does not make a determination of misconduct" and that the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors". Institutions vary in how closely they follow that, which is the real source of risk.
Only instructors and administrators see it
Turnitin's FAQ states plainly that "only instructors and administrators are able to see the indicator". Unless your institution enables student self-check, you cannot see your own AI score before submitting, and you cannot verify any vendor claim about clearing it either.
The trade-off nobody quotes
Turnitin says its own score is a floor
Every detector sets a dial between two kinds of error. Push false positives down and you miss more real AI. Push detection up and you start accusing people who wrote their own work. Turnitin picked the first, said so in writing, and published the size of the cost:
"In order to maintain this low rate of 1% for false positives, there is a chance that we might miss some AI written text in a document. We're comfortable with that since we do not want to incorrectly highlight human-written text as AI-written. For 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."
Read that as an instructor and it means the number on the report is the minimum, not the estimate. Read it as a student and it means the reverse of what you would hope: a low score is not evidence that a paper is clean, because the model is tuned to under-call rather than over-call.
This single paragraph is more disclosure than any other detector in this category offers, and it is the strongest argument that Turnitin is engineered seriously. It is also the reason no vendor, ours included, can tell you what score a given piece of writing will get.
Hard limits
What Turnitin's AI detector will not score at all
Accuracy questions usually assume the detector runs. Often it does not, and these limits come straight from Turnitin's file requirements and its definition of qualifying text.
| Constraint | Turnitin's rule | What happens in practice |
|---|---|---|
| Minimum length | At least 300 words of prose | Shorter work generates no AI score at all |
| Maximum length | No more than 30,000 words | Longer theses will not process |
| Languages | English, Spanish, Japanese | Anything else returns an error state, not a score |
| File types | .docx, .pdf, .txt, .rtf, under 100 MB | Other formats are rejected before analysis |
| Non-prose text | Not reliably scored | Poetry, scripts, code, bullet points, tables and annotated bibliographies are excluded from qualifying text |
| Similarity score | Completely separate | The AI percentage is "different from and independent of the similarity score", and AI highlights never appear in the Similarity Report |
The non-prose exclusion matters more than it looks. Turnitin only analyzes what it calls qualifying text, meaning full sentences inside paragraphs. A document heavy in bullet points, tables or code can return a percentage that describes a small fraction of the words on the page, which is why the AI percentage and the visible highlights sometimes disagree.
Current as of August 2026
The model changed in July 2026, so most accuracy articles are stale
Turnitin's FAQ records the change in one line: in July 2026 it "updated our model architecture to consolidate a multi-model ensemble into a single model", and states the update maintains the same sub-1% false positive rate. That is a month old at the time of writing. Every independent test anyone can cite predates it by two years or more.
The model list is worth reading too, because it is unusually current. Turnitin says its English detector covers output from GPT-5.4 and GPT-5.4-pro, Claude Sonnet-4.6, Claude Opus-4.5, Gemini-3.1-pro, Grok-4.1, LLaMA-4-Maverick, Deepseek-v3.2 and about twenty others, plus tools built on top of them. Anyone telling you a particular frontier model is invisible to Turnitin is contradicting Turnitin's own published list. We looked at that question in detail in our guide to which AI is least detectable.
One more change is easy to miss. Turnitin Clarity now ships a Writing Report that records how a document was written rather than guessing from the finished text: pasting activity, a playback timeline of the document developing, and the student's AI chat interactions when the AI assistant is enabled. That is a provenance feature, and no rewriting tool of any kind can affect it. It only exists inside Clarity assignments, and students see it only when the instructor turns it on.
The part that argues against us
Turnitin says it detects humanizers, and names one
We sell a rewriting tool, so we will state this plainly rather than leave you to find it. Turnitin's documentation says its English model flags text that was likely AI generated and then "further modified by an AI bypasser, AI-paraphrasing tool or AI word spinner, such as Quillbot". Its Spanish and Japanese detectors do not include that capability, but the English one does, and English is what most of this audience submits.
What follows from that is a limit on what any humanizer can honestly promise. Nobody selling one, us included, can see your institution's Turnitin configuration, run your document through it, or observe the score. Any page offering a guaranteed pass is making a claim it has no mechanism to verify. We have written separately about what an AI humanizer can and cannot actually do.
The useful version of the product is narrower and real. AI drafts share a flat, evenly-paced register: uniform sentence length, predictable connectives, no idiosyncrasy. Rewriting that register produces text that reads like a person, which is worth doing on its own terms. It is not a guarantee against a detector, and provenance features like Clarity are outside its reach entirely.
Excluded on purpose
The Turnitin accuracy numbers we will not repeat
Search this question and you will meet precise-looking figures on page after page: a 4.2% false positive rate, a 91% detection rate, a 98% accuracy claim, detection dropping to a tidy "60 to 85%" after editing. We went looking for the studies behind them. None of the pages carrying those numbers publishes a sample size, a date, a method or a prompt set, and almost every one of them belongs to a company selling either a humanizer or a competing detector. The numbers cite each other rather than any measurement.
A two-decimal figure with no method behind it is not more precise than an honest range. It is just harder to check. So the numbers on this page are limited to the ones with a traceable origin, each labelled with who produced it and what it does not cover, which is the same standard we apply on our AI detector false positive rate comparison across every major tool.
Across real-world text, the defensible summary for the category as a whole is roughly 85 to 90% accuracy with 8 to 15% false positives, and materially worse on writing by non-native English speakers. Turnitin's own figures sit better than that, and are also unaudited. Both things are true.
Frequently asked
Questions people ask about Turnitin AI detection accuracy
How accurate is Turnitin AI detector?
Turnitin publishes a false positive rate of under 1%, but only for documents it has already scored above 20% AI. It validates that figure against more than 700,000 pre-ChatGPT academic papers before each model release. Independent testing is thin: a Washington Post test in April 2023 found roughly 50% accuracy on a small sample, and Inside Higher Ed reported it missing about 15% of AI text in February 2024. Both predate the July 2026 model.
Does Turnitin AI detection have false positives?
Yes, and Turnitin says so. Its documentation states that false positives "are a possibility in AI models" and that its own testing found a higher incidence of them between 0 and 19 percent, which is why it hides every reading in that band behind an asterisk. It also names the writing most likely to be wrongly flagged: text with little structural variation, text that repeats itself, and text paraphrased without developing new ideas.
Can students see their Turnitin AI score?
Usually not. Turnitin's FAQ states that "only instructors and administrators are able to see the indicator". Some institutions enable a student-visible self-check or the newer Clarity Writing Report, but that is a per-assignment setting an instructor has to turn on. In most cases the first person to see your AI score is the person grading you.
What percentage of AI does Turnitin flag?
Anything from 20% upward is shown as a number. Between 1% and 19% Turnitin displays an asterisk and attributes no percentage, and 0% means no qualifying text was identified as AI. There is no universal pass mark, because the threshold for action is set by your institution rather than by Turnitin.
Does Turnitin detect AI humanizers and paraphrasing tools?
Turnitin says it does. Its documentation states the model detects text "likely AI generated" that "may have been further modified by an AI bypasser, AI-paraphrasing tool or AI word spinner, such as Quillbot". That capability is English-only; the Spanish and Japanese detectors do not include it. Any vendor promising you a guaranteed clear score, including any humanizer, is promising something it cannot verify.
How does Turnitin detect AI writing?
It splits a submission into overlapping segments of sentences, scores each segment between 0 and 1 for the probability of being AI-generated, lets each qualifying sentence inherit its segment score, pools sentences that appear in more than one segment, then aggregates those into one document score. We break the process down step by step in our guide to how Turnitin detects AI.
Is Turnitin AI detection accurate in 2026?
The model changed in July 2026. Turnitin consolidated what had been a multi-model ensemble into a single model and says the change maintains the same sub-1% false positive rate. No independent test of that version has been published yet, so every accuracy figure you will read about the current Turnitin, including the ones on vendor blogs, is either Turnitin's own or is measuring an older model.
Keep reading
Related guides on detector accuracy
Turnitin AI checker: can you check your score?
Why no public Turnitin AI checker exists, what the lookalike sites actually are, and what Turnitin shows students instead.
How does Turnitin detect AI writing?
The segment-and-pool mechanism behind the percentage, described step by step from Turnitin's own documentation.
What AI percentage is allowed in Turnitin?
Why there is no universal pass mark, and what the asterisk in the 1 to 19 percent band actually means.
AI detector false positive rates compared
Every major detector's published rate beside the independent measurements, and the figures with no traceable source.
Best AI humanizer for Turnitin
What rewriting can realistically change before a Turnitin submission, and what it cannot touch.
How to prove you did not use AI
The provenance evidence that actually holds up when a detector flags work you wrote yourself.
Originality.ai vs Turnitin
Two detectors built for different buyers, compared on what each one publishes about itself.
Write drafts that read like a person wrote them
No detector vendor can promise you a score, and neither can we. What we can do is take the flat, uniform register out of an AI draft so it reads the way you write.
Plans planned from $9/month billed yearly · Your runs saved to your history, delete any time