Short answer: yes, Turnitin can detect Claude. Turnitin's detector is built to flag text from large language models in general, and Turnitin's guidance lists Claude among the models it covers. It does not publish a Claude-specific detection rate, and detection gets weaker as text is edited or rewritten. Claude is one of the strongest writing models available, but it is not invisible to a detector, and treating it as a way to slip past Turnitin is a misunderstanding of how the check actually works.
Last verified
This guide explains why Turnitin catches Claude even though it was not built around that one model, how accurate the check really is, where it fails, and what an AI score does and does not prove. If you wrote something yourself and you are worried about a false flag, you can check your own work first and see the same signals a grader would see.
Turnitin does not look for "Claude" or "ChatGPT" by name. It looks for the statistical fingerprint that almost all large language models leave behind, regardless of which company made them. Because Claude produces the same kind of smooth, evenly weighted prose that other models do, it carries that fingerprint and gets flagged.
Two facts set the frame for everything below:
Keep both of those in mind. They explain both the catches and the false positives.
Turnitin has said its AI model was first trained on output from GPT-3 and GPT-3.5, and it has since expanded coverage to newer models, with its AI writing detection capabilities FAQ listing OpenAI GPT, Google Gemini and Anthropic Claude models among them. But the deeper reason it catches Claude is not that it memorized Claude's style. It is that all of these models share a way of writing that differs from how people write.
Further reading: Does Turnitin Detect ChatGPT?, on Substack.
Human writing is uneven. Some sentences run long and looping, others are short. Word choices wander, rhythm shifts, and a writer will occasionally make an odd or inefficient choice. Language models, including Claude, tend to pick the most statistically likely next word and produce text that is smoother and more uniform than human prose. Detectors measure that evenness through two properties researchers call perplexity (how predictable each word is) and burstiness (how much sentence-to-sentence variety there is). Machine text scores low on both, and that low score is the tell.
Claude is good enough that its output often reads beautifully. That polish is exactly the problem from a detection standpoint. The cleaner and more confident the prose, the more it can look like the even, predictable pattern the detector is tuned to flag. This is the same reason Turnitin catches ChatGPT and Gemini: the detector is reading the shape of the writing, and the shape is similar across all three models.
No. On 14 August 2026 Anthropic announced that future Claude models will write text containing a statistical watermark, and that it is working to add one to models launched before 2 August 2026. Only Anthropic's key can check it, and Anthropic says plainly that "AI detection software uses a different method, because the companies that provide it don't have our key." A Turnitin AI score is therefore not a watermark check.
| Turnitin AI score | Claude's watermark | |
|---|---|---|
| What it reads | Patterns in the writing that look machine-made | A hidden statistical pattern Claude adds as it writes |
| Who can check it | Institutions that have Turnitin's AI detection turned on | Only Anthropic's key; its detection API is in private preview for eligible organisations, as EU law requires |
| The question it answers | How likely is it that this text is AI-generated, by any model? | "What is the likelihood this was partly written by Claude?" It cannot confirm human writing or spot another AI |
| Names Claude? | No | Only in the sense above: Claude or not Claude |
| After editing | Weaker as text is rewritten | "Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will." |
Anthropic also says the watermark does not affect the quality of Claude's output. Sources: Anthropic's announcement, checked 2 November 2026.
Here is where the marketing and the reality part ways, so it is worth being precise.
Turnitin does not publish a detection rate for Claude specifically. We could not find a Turnitin source for any per-model detection rate, so we do not quote one. What Turnitin does publish is a false-positive figure: its documentation reports a document-level false-positive rate under 1% for documents it scores at 20% AI or more, and roughly 4% across the full distribution, with no percentage shown below 20% (only an asterisk). Those are vendor numbers from Turnitin's own evaluation set. Our Turnitin accuracy review goes through them in detail.
Independent research on detectors as a class is less flattering. Weber-Wulff et al. (2023) tested 14 detection tools, Turnitin among them, and concluded they were "neither accurate nor reliable", with a bias toward classifying machine output as human-written, especially once it had been edited or paraphrased. As soon as text is edited, paraphrased, or rewritten in a person's own voice, detection weakens, because the human changes break up the machine evenness the detector relies on.
So the honest statement is layered:
Anyone who tells you a detector is a perfect lie-detector for Claude is overselling it.
A high catch rate on machine text does not mean a clean record on human text. Detectors flag genuine human writing as AI sometimes, and that error is not random. It falls hardest on a few groups.
Non-native English writers are flagged more often, because writing carefully in a second language can produce the same even, measured phrasing that machine text has. A Stanford study (Liang et al., 2023) found that seven GPT detectors misclassified an average of 61.22% of 91 TOEFL essays by non-native speakers as AI-generated, far more than native-speaker essays. Turnitin was not one of the seven, so that is not a Turnitin figure, but the mechanism applies to detectors as a class. These fairness concerns were serious enough that some universities turned Turnitin's AI detection feature off rather than risk wrongly accusing students; Vanderbilt disabled it in August 2023.
This is why Turnitin itself states that its AI score "should not be used as the sole basis for adverse actions against a student." The company that builds the tool is telling you the number is a starting point for a conversation, not a confession. If you have been flagged for writing you actually wrote, our guide on how to prove you didn't use AI covers the evidence that helps most (version history, drafts, a second detector check) and how to present it.
The number Turnitin returns is one of the most misread figures in education, so it is worth slowing down.
An AI score is a percentage estimate of how much of the document the model believes is AI-generated, expressed as a probability. It is not a measurement of fact. A "40% AI" result does not mean exactly 40% of the paragraphs were written by Claude. It means the model is partly confident about part of the text and unsure about the rest. A "0%" does not prove a human wrote every word, and a "100%" is a strong signal but still an estimate, not a recording of what happened.
Three things follow:
Detection is strongest on unedited text, so it is fair to ask what editing does. The mechanism is straightforward, and understanding it is not the same as a how-to for evasion.
The detector keys on uniformity. When a person reads machine output, rewrites sentences in their own voice, adds their own examples, breaks the even rhythm, and corrects the places where the model said something generic or slightly wrong, they introduce the human variance that detectors are built to notice the absence of. The more the text becomes genuinely the writer's own work, the less it looks like raw model output, which is the entire point: text that a person has truly made their own is, increasingly, their own.
That said, Turnitin closed one common shortcut. Since August 2025 it also flags text that was AI-generated and then run through a paraphraser or "humanizer" to disguise it. So mechanically spinning Claude output through a bypass tool is not a reliable escape; the system is specifically tuned to catch that pattern now. The reliable path was never a tool. It was doing the thinking and the writing yourself.
If you used Claude to brainstorm or draft and then rewrote the piece in your own words, or if you wrote the whole thing yourself and you are simply anxious, the most useful move is to see what a detector sees before anyone else does.
TextSight's free AI Detector gives you an overall AI probability score, and Starter adds sentence-level highlights, so you can find the exact lines that read as machine-generated and revise them honestly in your own voice. We are also clear about the limits, the same limits that apply to Turnitin: the result is a probability, not a verdict, it can produce false positives, and it is less certain on edited or second-language writing. Used the right way, before submission, on your own work, as a self-check, it is genuinely helpful. It is not a way to game a grade, and we do not build it to be one.
If you are on the other side of the desk, our AI detector for professors and faculty page covers sentence-level review of student work, and why draft history and a conversation with the student matter more than any single score.
TextSight is an AI-detection and writing-trust tool. We help you check and improve your own work. We do not help anyone disguise AI writing or evade an institution's detection. Detection scores, ours and Turnitin's alike, are probabilities and can be wrong, especially on edited or second-language writing, so treat any score as a guide for revision and a prompt for review, not as proof.
TextSight's free tier gives you daily scans with an overall AI score, and Starter adds sentence-level highlights so you can see which lines carry the AI signal. Every score is a probability, not proof, and we say so plainly.