The short answer is that no outside tool has been shown to reproduce a Turnitin score, and a tool advertising that it does is making a claim nobody can verify. The more useful answer is that matching the number was never the goal worth pursuing. What you actually want before submitting is a second reading that shows you which passages a detector treats as machine-like, so you can look at them yourself. That is a property several tools offer and many do not. Turnitin publishes its file requirements and its reporting rules, but not the training corpus or the decision boundary that would let anyone calibrate a competing tool against it. This article sets out the criteria that separate them, gives a self-check worth running, and is honest about what a self-check cannot do for you.
Why no outside tool reproduces the score
Turnitin publishes a fair amount about its system and withholds the part that would let anyone replicate it.
What is public: the file requirements, the report structure, the supported languages, and the policy of not attributing a score between 1 and 19 percent. Turnitin also describes the sample it trained on in general terms, including an effort to represent second-language learners and users from non-English-speaking countries. Turnitin's published requirements and report categories cover this in detail.
What is not public: the corpus itself, the model architecture, the features it uses, and the internal decision boundary that turns a continuous output into a reported percentage. Without those, an outside tool cannot be calibrated against it, and no independent test has demonstrated that any tool tracks Turnitin's output closely across varied text.
So treat score matching as unverifiable rather than merely difficult. What a second tool gives you is narrower and still useful. A passage it flags may deserve a second look, not because the flag establishes that anything is wrong with the writing, but because it shows how that particular model read the text.
The criteria that actually separate these tools
Since matching the percentage is off the table, here is what to compare instead.
| What to check | Why it matters |
|---|---|
| Sentence-level highlights | Turnitin highlights passages, not just a total. A tool that returns one number gives you nothing to inspect or revise. |
| Long-form support | Turnitin publishes its file requirements at 300 words minimum. A tool that returns confident verdicts on two sentences is not matching your real use case. |
| A readable privacy policy | You are pasting unpublished academic work. Retention and training practices vary by provider. Check whether the policy states how long submissions are kept, whether they are used for model training, and whether third-party processors receive the text. |
| Mixed-text behavior | Real drafts contain human and machine sentences together. This is the hardest case for every system, and a tool that handles it as an all-or-nothing verdict is oversimplifying. |
| Language coverage | Most detectors were built and validated on English. If you write in another language, ask what evidence exists for that language specifically. |
| Stated limitations | A tool that publishes what it cannot do is easier to use correctly than one that publishes only an accuracy figure. |
The practical conclusion is that the closest tool is not the one promising the same percentage. It is the one that shows you sentences, handles long-form text, tells you what it does with your data, and is explicit about where it stops being reliable.
A self-check worth running
Ten minutes, before you submit, not after you are accused.
One. Run the complete draft through a detector and open the sentence view rather than the summary. Unfox is an AI detector that reports at this level, and several other tools do too. What you are looking for is whether flagged passages cluster in one section or scatter across the document.
Two. Read each flagged sentence and ask whether it accurately reflects your reasoning, and whether you could explain why it is phrased that way. A sentence you cannot explain, or one whose wording sits oddly against the rest of the draft, deserves a closer look regardless of how it was produced. That question is more useful than asking whether it sounds like you, because human sentences can read stiffly and machine sentences can read casually.
Three. Run it again somewhere else. Agreement between tools narrows the list of passages worth a second look. Disagreement narrows it differently, by showing you where the signal is weak rather than strong, and detectors can disagree with each other on the same text more than most people expect.
Four. Save the draft, the report, and the date together. That pairing costs nothing now and is the only version of it that exists later.
If you are checking documents on a schedule rather than one paper once, the detection API covers that case.
What a self-check cannot do
It cannot predict your institution's result. Different model, different threshold, different qualifying-text rules. A clean report elsewhere is not a forecast.
It cannot clear you after an accusation. A second detector report is another probability estimate and carries the same weakness as the first one. Drafts, version history, and being able to talk through your reasoning are what address the question actually being asked, which what to do if you have already been flagged covers in full.
And it cannot fix the underlying reliability problem in this category. Research on detector bias against non-native English writers published in Patterns found that widely used detectors flagged human-written essays by non-native speakers at a far higher rate than comparable native-speaker samples. Turnitin says it accounted for second-language writers when assembling its training sample, which is a reasonable response to the finding and not a refutation of it.
Use a self-check to find passages worth reviewing. It is not a verdict, and no detector output should be treated as one.
FAQ
Which AI detector is closest to Turnitin?
No outside tool has been shown to reproduce Turnitin's score, and the training data and decision threshold that would make that possible are not public. What some tools share is the sentence-level view, which lets you inspect the kind of passages an instructor's report would highlight.
Can I run my own paper through Turnitin before submitting?
Usually not directly, since most institutions route Turnitin through the assignment system and send the AI score to instructors rather than students. Some schools enable a draft submission option, so it is worth asking, because a check inside your own course beats any outside estimate.
Is a free AI detector good enough?
For locating passages worth reviewing, yes. For deciding whether you are in trouble, no, and that applies to paid tools too. Price does not change what these systems output, which is a probability rather than a determination.
Will another detector give me the same score as my school's?
Very likely not, and matching numbers is the wrong target. Different training data and thresholds produce different percentages from identical text. What is worth comparing is which passages each tool highlights.
Is it safe to paste my unpublished essay into a detector?
It depends on the tool, and this is worth checking rather than assuming. Retention and training practices vary by provider, and the terms are usually in the privacy policy rather than on the marketing page. Look for how long submissions are kept, whether they are used for training, and whether third parties process the text.
How long does my text need to be?
Longer than most people assume. Turnitin requires 300 words of qualifying prose and notes that accuracy improves with more text. Short passages produce unstable results across this category, so treat a confident verdict on a paragraph with caution.
