Yes to the first, and a more specific yes to the second. Turnitin is designed to detect writing produced by major generative AI systems, and its AI writing report includes a separate category for text it believes was AI-generated and then reworded by an AI paraphrasing tool, with QuillBot named in the documentation as an example. What most articles on this question leave out is the set of conditions attached. The report needs at least 300 words of prose in a long-form format, it stops at 30,000 words, and it runs only on English, Spanish, and Japanese submissions. Miss any one of those and no AI score is produced at all, which catches out more short assignments than most students expect. This article covers what Turnitin publishes about its own system, which claims it does not make, and where the report stops being informative.
The requirements most articles skip
Before any of the detection questions matter, a submission has to qualify. Turnitin publishes these thresholds, and a document that misses one of them produces no AI score at all rather than a low one.
| Requirement | Current value |
|---|---|
| Minimum qualifying text | 300 words of prose in a long-form format |
| Maximum qualifying text | 30,000 words |
| Supported languages | English, Spanish, Japanese |
| Accepted file types | .docx, .pdf, .txt, .rtf |
| File size | Under 100 MB |
| Results between 1 and 19 percent | Displayed as *%; no exact percentage and no highlights |
| Submission that does not qualify | No percentage is displayed and no AI writing report is generated |
Two of these matter more than the rest. Short assignments, discussion posts, and supplemental application essays frequently fall under 300 words, which means the AI report never runs on them. And Turnitin notes that accuracy improves with more text, so a submission just over the threshold produces a less reliable score than a long paper does.
The language list is also narrower than most people assume, and Turnitin describes its Spanish and Japanese models as different models from the English one rather than the same system applied to another language.
Two numbers, and you only see one
A Turnitin submission can produce two separate results, and confusing them is the most common misunderstanding students bring to this topic.
The similarity score measures how much of your text matches material in Turnitin's database. It is a text-matching result rather than a determination of plagiarism, since quotations, references, and common phrasing all produce matches, and it has nothing to do with the AI writing score. Whether you can see it yourself depends on your institution's and instructor's settings.
The AI writing score is a different report that estimates what proportion of your qualifying text the model judged likely to be AI-generated. Student access varies by institutional setup, but this report is commonly instructor-facing rather than automatically visible to you. That asymmetry deserves naming. The number shaping how your paper gets read is often one you never see.
Turnitin also does not attribute a score or highlights for results in the 1 to 19 percent range, on the grounds that false positives concentrate there. In that band the indicator displays an asterisk, shown as *%, with no exact percentage and no highlighted passages.
What the AI writing report actually contains
The report breaks the overall percentage into two interactive categories, each highlighted separately in the submission.
The first is AI-generated text, meaning qualifying passages the model judged likely to have come from a large language model. Turnitin notes that this category can include text modified by tools designed to defeat detection.
The second is AI-generated text that was AI-paraphrased, meaning passages the model judged likely to have been generated by AI and then revised with a paraphrasing tool or word spinner. Turnitin's documentation gives QuillBot as its example, and this category is highlighted in a different color.
Turnitin publishes the report structure. It does not publish the model architecture, the feature set, or the weighting behind either category, so descriptions of exactly how the system reaches a conclusion, including confident ones you will find on other sites, are inference rather than documentation.
The AI-paraphrased category is narrower than it sounds
This distinction gets lost constantly, and it changes what the category means for you.
Turnitin describes the category as text that was likely AI-generated and then likely modified by a paraphrasing tool. It is a two-step pattern, not a rule that any text touched by QuillBot lands there. Running your own writing through a paraphraser is not the scenario the category describes.
Treating this as a safe distinction would still be a mistake, because you do not control which category a model assigns your text to, and heavy rewriting changes prose in ways that are hard to predict. What a paraphrasing tool changes covers the practical consequences.
One coverage note. Turnitin's paraphrasing and bypasser detection capabilities are documented for its English model. The Spanish and Japanese models are described separately and do not carry the same capabilities.
Where ChatGPT, Claude, and Gemini actually stand
Turnitin states that its detection capabilities target writing from major large language models and that its models are updated as new systems appear. It has published which models specific language versions were trained against, and it updates that list over time.
What Turnitin does not publish is detection performance broken out by model. There is no official figure for how the English model performs on ChatGPT versus Claude versus Gemini, so treating them as equivalent is an assumption rather than a fact, and so is treating any one of them as safer.
The same limit applies to model families that appear less often in public evaluations, including DeepSeek, Qwen, Kimi, and Doubao. Published detector benchmarks tend to concentrate on a small number of widely used systems, and no commercial vendor discloses its full training corpus. When a detector is evaluated on a distribution unlike its training data, performance becomes harder to predict, in either direction. Anyone telling you confidently that a particular model slips past a particular detector is guessing.
What the report does not tell you
The standard AI writing report analyzes a submitted document. It does not establish how that document was produced, and Turnitin's own guidance is that the AI score should not be used as the sole basis for action against a student.
One common belief now needs a caveat, though. If your institution uses Turnitin Clarity and you write inside its composition space, Turnitin Clarity records paste events and a writing timeline alongside writing time and session data. Clarity only tracks work done in its own editor, so drafting elsewhere and uploading a file produces no timeline, but a large paste into the Clarity workspace is logged as an event. Whether any of this applies to you depends entirely on what your institution has enabled.
Two more limits. Basic spelling and punctuation correction is unlikely to move an AI result much, since it changes very little text, but generative rewrite, clarity, and tone features in the same products can rewrite whole sentences and should be treated as a different category of tool. And a report is a probability estimate, not a finding about a person, which is the subject of why human writing gets flagged. Unfox is an AI detector, and that caution applies to every tool in this category, ours included.
Checking before you submit
Read your own draft through a detector before you submit, with one adjustment. Look at the sentence-level output rather than the headline percentage. The number tells you very little. The specific passages tell you where a reader might misread you.
Keep your drafts, outlines, and version history as you go. In a dispute, process is the material that speaks to the actual question, and it has to exist before you need it.
For choosing which tool to use, choosing an external checker covers what to look for. For checking many documents at once rather than a single paper, batch detection handles the volume case.
FAQ
Does Turnitin detect ChatGPT?
Turnitin is designed to detect writing from major large language models and updates its models as new systems appear. It does not publish detection rates broken out by model, so no reliable figure exists for ChatGPT specifically. The report also requires at least 300 words of qualifying prose before it runs at all.
Can Turnitin detect QuillBot?
Turnitin has a category for text it judges likely AI-generated and then likely reworded with an AI paraphrasing tool, and its documentation names QuillBot as an example. Note the two-step pattern. The category describes AI text that was then paraphrased, not any text a paraphraser touched.
What is the minimum word count for Turnitin AI detection?
300 words of prose in a long-form format, with a maximum of 30,000 words. Submissions below the minimum do not generate an AI writing score. Turnitin also notes that accuracy improves with longer text, so a document just above the threshold gives a less reliable result.
Which languages does Turnitin AI detection support?
English, Spanish, and Japanese. Turnitin describes the Spanish and Japanese systems as separate models rather than the English model applied to other languages, and the paraphrasing and bypasser detection capabilities are documented for the English model.
Can Turnitin see my drafts or copy-paste history?
Not from a standard file upload, which is analyzed as a finished document. If your institution uses Turnitin Clarity and you write inside its composition space, paste events, writing time, and a playback timeline are recorded. Work drafted outside Clarity and uploaded does not generate that report.
What AI score is considered safe?
There is no published safe threshold. Turnitin does not attribute a score in the 1 to 19 percent range and advises that the result should not be the sole basis for action against a student. What happens next is set by your institution's policy rather than by the number.
