CiteOwl

Blog/Using AI honestly

Can Turnitin detect Claude? What the published list says

The CiteOwl Team· Updated 24 August 2026· 11 min read

Yes, and Turnitin will tell you which Claude it means: its published coverage list names seven Claude versions outright, from Claude 3 Haiku up to Claude Sonnet 4.6, plus any tool built on them. But the list stops there. No Claude released since February 2026 appears on it, and Turnitin has never said what happens to a model it has not named. Running alongside it is a second system nobody at your university can use, because Anthropic now marks Claude's text with a watermark no detector can read. This guide reads both against their own documentation and shows why neither can tell "Claude wrote this" from "Claude edited this".

Almost every page answering this question hands you a percentage, and none of them can show where it came from. What can be shown is Turnitin's own documentation, Anthropic's own documentation, and the two peer-reviewed studies that ran Claude output through Turnitin. That is what this page is built from. For how the AI indicator behaves in general, our guide on whether Turnitin can detect ChatGPT covers the mechanism and the false positives.

Turnitin names the Claude versions it covers

Turnitin's AI writing detection is a separate feature from the similarity score, and it does not compare your essay against a library of chatbot output. It runs your sentences through a classifier. In its AI writing detection capabilities FAQ, accessed 24 August 2026, it publishes something unusual for a detection vendor: a list of the specific models its English detector picks up, 32 of them, each with a release date. Seven are Claude.

The sentence closes with a clause worth reading twice. Coverage extends to "tools based on these LLMs as well". If your writing assistant runs on Claude underneath, which many do without saying so on the marketing page, Turnitin's claim reaches it.

What the list does not do is give the report a name to print. The detector never says "this looks like Claude". It scores each sentence for how likely it is to be machine-generated at all, so Claude text and ChatGPT text come back as the same kind of percentage. Turnitin is also blunt about how that score is produced: its model, it says, "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions". Most of the advice about Claude's supposedly higher burstiness is arguing about a dial Turnitin says it does not turn.

Every Claude since February 2026 is missing from it

Set the list against Anthropic's own model documentation and a gap opens. The newest Claude Turnitin names is Sonnet 4.6, from February 2026. Anthropic's current docs describe Opus 4.6, Opus 4.7 and Opus 4.8, Sonnet 5, Opus 5, and Claude Fable 5, available since 9 June 2026 and described there as Anthropic's most capable widely released model. None of them appear on Turnitin's list.

The list is patchy inside the Claude line too, not just behind it. It jumps from Sonnet 3.7 straight to Sonnet 4.5, skipping the Sonnet 4 and Opus 4 generation. And it disagrees with itself: the same page prints the list twice, dating Sonnet 4.6 to 2026-02 in one place and 2026-05 in the other. That is a hand-maintained document, not a live registry.

Here is where the humanizer sites want you to draw a conclusion, so be careful. An absence from that list is not a statement that the model gets through. Turnitin does not claim its detector generalises to models it has not named, and it does not claim the opposite either. It does say the classifier keys on statistical patterns rather than model identity, and that in July 2026 it consolidated a multi-model ensemble into a single model while holding its false positive rate under 1%. A detector built that way is not promised to fail on an unlisted model. Nobody, Turnitin included, publishes a number for the version you used.

One practical note that follows from the release notes: model updates never rescore old work. A submission is only re-examined if it is submitted again.

What independent testing actually found

Two peer-reviewed studies have put Claude output through Turnitin. Both are worth knowing, and both stop short of the number people want.

The first is a 2024 study in the International Journal of Educational Technology in Higher Education, which generated text with GPT-4, Bard and Claude 2, then ran it through six detectors including Turnitin. At baseline, Turnitin caught 61% of AI text, second only to Copyleaks at 64.8%, with Bard the most detectable generator at 76.9% of its outputs identified.

The interesting part is what happened under editing. Averaged across detectors accuracy fell 17.4%, but the drop was uneven by generator: Bard lost 38.8%, Claude 2 lost 8% and GPT-4 lost 7.6%. The authors read that as Claude 2 already sitting close to human writing patterns before anyone touched it. Turnitin took the largest hit of any tool tested, 42.1%, falling from second place to fifth.

The second is a February 2026 study in the International Journal for Educational Integrity, which built a balanced set of 192 texts from student writing, professional writing, and output from GPT-4.1 and Claude 3 Opus, then tested Turnitin and Originality. Turnitin's macro-average accuracy came out at 0.61, with recall at 0.51. Subject matter mattered more than most people expect: 0.86 on humanities writing against 0.51 on scientific writing, a gap the authors found statistically significant. An essay on urban heat policy and the same argument in a methods-heavy register are not the same problem for this tool.

Now the caveat neither study lets you skip. Both pool their results. The 2024 paper averages across six detectors, so it never publishes a Turnitin-on-Claude figure, and the Claude it tested was Claude 2, a 2023 model.

The 2026 paper reports Turnitin alone but pools Claude 3 Opus with GPT-4.1. Meanwhile the most recent study dedicated to Turnitin, published in the Journal of Applied Learning and Teaching in January 2025, tested ChatGPT, Perplexity and Gemini and left Claude out altogether, reporting Turnitin holding a 100% AI score even through adversarial edits. That sits awkwardly beside the 42.1% collapse in the 2024 paper, which tells you how unsettled this literature is.

The percentages on this search are invented

Search this question and you will be handed numbers with two decimal places. We chased them. One widely cited page reports 150 samples producing detection rates of 89% for Opus, 87% for Sonnet and 83% for Haiku, with no method and no third party. Another cites a 2024 study in the journal Computers and Education, 500 human essays against 500 AI essays, 91% detection and a 4.2% false positive rate.

We searched that journal's catalogue. No such paper exists. The same sites claim Turnitin has confirmed its training data skews toward ChatGPT output. It has confirmed no such thing, and its documentation says only that early work focused on GPT-3 and GPT-3.5.

Those pages sell humanizers or essay-writing services, and a precise-sounding percentage is the product demo. If a number about Claude and Turnitin arrives without a method you can read, treat it as marketing, which is also the reasoning behind what humanizers actually do to a score.

Claude marks its own text, and Turnitin cannot read it

There is a second detection system here, and it belongs to Anthropic rather than to your university. Since August 2026 Claude has woven an imperceptible watermark into its own output, worldwide, across the API, Claude Code and the chat app alike. Nothing is added to the text and there are no hidden characters. As Anthropic explains in its account of how the watermarking works, published 14 August 2026, the method is a version of Google DeepMind's SynthID-Text: where several next words would be equally good, a secret key decides which one Claude picks.

Two facts matter more than the mechanism. The first is who can read it. Asked directly how its watermark differs from AI detection software, Anthropic answers that detection companies "don't have our key". Turnitin does not have it, and neither does any tool your department licenses. Anthropic says it will offer a detection API and is "in the process of working out the details", which is not a date. The mark exists and nobody grading essays can check it.

The second is what it would prove if they could. Anthropic states the mark carries no identifying information and cannot be traced to a person, organisation or chat, and its support documentation calls a detected mark a signal that content was processed by Claude, "not fully conclusive". We covered the whole marking landscape, Google and OpenAI included, in does AI text have a hidden watermark.

Using it for help is the case both systems handle worst

Most people asking this question did not paste an essay out of a chatbot. They wrote a draft and asked Claude to tighten it, or drafted two paragraphs with it and rewrote the rest. That middle ground is exactly where both systems are least informative.

Start with Turnitin. It suppresses low scores entirely: anything from 1% to 19% shows as an asterisk with no percentage, precisely because false positives are likeliest there. It admits its scores understate, giving the example that a document read as 50% AI "could contain as much as 65% AI writing". On short work it warns that in documents of a few hundred words "the prediction will be mostly 'all or nothing'", so a genuine mix can be flagged as entirely machine-written.

The February 2026 study found the same shape from outside: on texts built half from student writing and half from AI, Turnitin was weak on every metric. Our guide to what a Turnitin AI score actually means goes further into reading one.

Anthropic's watermark is no better at this, for the opposite reason. The mark attaches only to words Claude itself chooses, so when Claude proofreads your writing, Anthropic says, "there's very little (if anything) for the watermark to attach to". It is thin on factual passages too, where the next word is not a free choice. Light editing will not remove it; a full rewrite will.

The answer that does not depend on a list

Read the two systems together and the shape is clear. Turnitin's coverage list is real, published and checkable, and it also lags model releases by months and contradicts itself on a date. Anthropic's watermark is real and unreadable by anyone who would care. Neither separates the essay you wrote from the essay you were helped with, which is the only distinction your department is asking about.

What does not move is the work underneath. A reference to a paper that was never written stays wrong through any amount of rewriting, and a marker who clicks one dead DOI has something far more concrete than a score. That is where AI assistance genuinely goes wrong for students, and anyone can check it in a minute. We cover the mechanism in why AI makes up citations and the test itself in how to check if a citation is real. The other durable exposure is an argument you cannot explain.

We should say where we stand, because the question lands on us too. CiteOwl's agent drafts with a frontier model run by a third party, so what that provider embeds is the provider's decision, not ours. We add no mark of our own, and we will not build a tool for removing anyone else's. What we do instead is the reason the product is shaped the way it is: every sentence the agent proposes arrives as a change you review beside your own text, and every claim carries the source it came from with the quote that supports it.

That standard survives any release note. Not "is my model on the list", but "can I stand behind every sentence and every source". Get that right and the list becomes trivia. If you are weighing the two assistants on research quality rather than detectability, our comparison of ChatGPT and Claude for research papers is the more useful question.