Does Claude Watermark AI Text? Can You Detect It? (Not Yet)

Last verified: 18 August 2026. Anthropic has confirmed new Claude models watermark their text. Anthropic has not yet released a way for anyone outside the company to check for that watermark. This page will be updated the moment that changes.

The short answer. Any Claude model released on or after 2 August 2026 watermarks the text it generates, from launch. Models that already existed before that date don't have it yet, Anthropic says that work is in progress. Either way, no you cannot detect that watermark yourself right now, and neither can Turnitin, GPTZero, Pangram, Originality.ai or any other AI detector, including ours. Anthropic holds the only key. It has said it plans to release a detection API, but as of this writing that API is not live. Every "can AI detectors catch Claude's watermark" claim you read today is either wrong or getting ahead of the facts.

What Anthropic actually changed

From 2 August 2026, every new Claude model embeds a statistical watermark directly into the text it generates. Anthropic's own explanation describes an approach adapted from Google DeepMind's SynthID-Text method: instead of adding hidden characters or anything a reader could notice, the watermark works by subtly biasing which words Claude picks at each step, according to a pattern only Anthropic's key can decode. No single word choice looks unusual. Across enough text, the pattern becomes statistically detectable, to someone who holds the key.

The change covers Claude wherever it is offered: claude.ai, the Claude Platform API, Claude Code, Claude Cowork, Claude Tag, and Claude accessed through AWS, Google Cloud or Microsoft Foundry. It applies globally, not only to European users, even though the immediate trigger was the EU AI Act's Article 50 transparency requirement, which took effect on 2 August 2026. Anthropic says older models will get watermarking added "in the coming months"; it is not retroactive yet.

Files are handled differently. When Claude produces a supported file type, such as a .png, .jpg or .svg, it attaches signed provenance metadata under the C2PA standard, the same open standard used across the AI industry for image and video provenance, rather than a text watermark.

What a Claude watermark can actually prove

Not authorship, and not certainty. Anthropic's own framing is that a detected watermark answers one narrow question: "what is the likelihood this was partly written by Claude?" It is a probability, not a verdict, and it does not distinguish between Claude writing a passage from scratch, Claude heavily editing a person's draft, or Claude proofreading text that is almost entirely someone else's. Anthropic is explicit that in that last case, the watermark barely attaches because nearly all the words are still the human author's.

A translation produced by Claude is watermarked in full, because every word in the output was Claude's choice. That is one of the few cases where the signal is strong.

What it does not prove

• It cannot tell you the text wasn't written by a different AI model. It only speaks to Claude specifically.

• It cannot confirm full provenance. Watermarked text can be excerpted, combined with other material, or passed through further edits after Claude produced it, and the detection result cannot untangle that history.

• It does not work reliably on short passages. Anthropic states detection needs enough word choices to establish a statistical pattern, so brief answers carry a weak or absent signal.

• It is sparser on factual, constrained writing, where there are fewer word choices Claude can make without reducing accuracy.

• Absence of a detected watermark is not proof the text is human. It could be an older model that predates the 2 August 2026 rollout, heavily edited text, a short passage, or a platform the marking doesn't cover yet.

Can ordinary AI detectors see it? Not this way

This is the part getting confused in a lot of the coverage. Turnitin, GPTZero, Pangram, Originality.ai, Copyleaks and every detector like them, including ours, do not have access to Anthropic's watermark key, and none of them are reading Anthropic's statistical pattern. They work through an entirely separate method: scoring how predictable, uniform or "AI-shaped" a passage's word choices and sentence rhythm look, based on patterns learned across many AI and human writing samples. We cover that approach in how AI detection actually works, and it has nothing to do with Anthropic's key.

So the honest answer to "does Turnitin detect the Claude watermark" is no. Whatever Turnitin flags in Claude-generated text, it is flagging through its own probabilistic method, the same one it has always used, not through Anthropic's mechanism. A Turnitin or Pangram score on a Claude-written essay today tells you the same kind of thing it told you about ChatGPT text last year: a statistical estimate, not a watermark match.

Anthropic has said it plans to release a watermark detection API that would let third parties check for the mark. That is not available yet. Current third-party AI detectors do not have access to Anthropic's watermark detector or key, so their Claude detection results are based on other methods entirely, and that will remain true until Anthropic actually ships third-party access.

What happens if the text is edited?

Anthropic's own stated evidence, not independent testing on our part, since there is currently no way to test this independently: light editing probably will not remove the watermark completely, because most of the original word choices survive. A complete rewrite, where every word is replaced, will remove it, though at that point Anthropic itself notes it becomes arguable whether the result should still be called "AI-generated" at all. Paraphrasing and moderate rewriting sit somewhere between those two, and Anthropic has not published a precise threshold.

Does Claude Code watermark output?

Only partially. Anthropic's guidance says that wherever an exact output is required, meaning there is no real choice between alternatives, the watermark isn't applied, and code very often falls into that category, so it carries generally less watermarking than prose. Where there is an arbitrary choice between particular words or terms within the code, such as in comments, the watermark can still apply.

Can the watermark be removed?

Anthropic's own answer, at a conceptual level: heavy enough editing removes it, because the method depends on Claude's specific word choices surviving into the final text. We are not going to turn this into a how-to, and neither does Anthropic. A page built to help people defeat a provenance system is a different, worse thing than a page built to answer what the system can currently prove.

Claude watermarking vs AI detection

Anthropic's watermarkTraditional AI detectors (Turnitin, GPTZero, Pangram, ours)

How it worksStatistical bias in word choice, embedded during generation, decoded with Anthropic's keyScores how AI-typical a passage's patterns look, based on training data, no generation-time cooperation needed
Who can check itAnthropic only, today. A third-party API is planned, not liveAnyone, via the detector's own tool
What it can flagText generated by covered Claude models onlyText from any model the detector was trained to recognise, with varying accuracy
Survives editing?Light editing: probably. Full rewrite: noDegrades with editing and paraphrasing, well documented, see can ChatGPT be detected
Proof or probabilityProbability ("likelihood partly written by Claude"), not proofProbability, not proof
Available to check todayNo, key not publicYes, that is what these tools already do

What users should actually rely on

If you are asking whether Claude text can be caught right now, in a school, a hiring process or an editorial workflow, the answer runs through existing AI detectors, the same probabilistic tools already being used on ChatGPT and Gemini text, not through Anthropic's watermark. That mechanism exists, but nobody outside Anthropic can act on it yet. Treat any current claim that a tool "detects the Claude watermark" as false until Anthropic's own detection access is public, and treat this page as the place to check when that changes.

Is It AI? works the same way today as it did before Anthropic's announcement: it looks at how a passage reads, not at a provider's key, so it is unaffected by whether a watermark exists underneath. If you want to understand what the false-positive and false-negative trade-offs of that approach actually look like, we publish our own numbers in how often AI detectors flag human writing.

When will IsItAI test this properly?

The moment Anthropic's watermark detection API is public, we intend to be one of the first to run real tests against it, both with the caveat that the underlying method is not directly comparable to statistical detection, and with a published update to this page. Until then, anything claiming to test the Claude watermark independently is not doing what it says.

Frequently asked questions

Does Claude watermark text right now?
Yes, for any Claude model launched on or after 2 August 2026, across claude.ai, the API, Claude Code, Claude Cowork, Claude Tag, and AWS, Google Cloud and Microsoft Foundry access. Older models do not have it yet.

Can I detect the Claude watermark myself today?
No. Anthropic has not released the detection mechanism to anyone outside the company. It has said an API is planned, with no confirmed date as of 18 August 2026.

Does Turnitin detect the Claude watermark?
No. Turnitin's AI detection, like GPTZero, Pangram and every other current detector, uses its own statistical pattern analysis, unrelated to Anthropic's key. See does Turnitin detect ChatGPT for how that method works and what its actual limits are.

Does editing remove the Claude watermark?
Anthropic says light editing probably will not remove it, and a full rewrite will. There is no independently tested threshold in between, because outside testing isn't currently possible.

Does Claude Code watermark the code it writes?
Mostly not the code itself, which very often needs an exact output with no real word-choice available, so it carries generally less watermarking than prose, per Anthropic's own explanation. Where the code has an arbitrary choice between wordings, such as in comments, the watermark can still apply.

Will third-party AI detectors ever be able to check the watermark?
Anthropic has said it plans to offer a detection API, which would let third parties check. It has not said detection access will never be available, and it has not shipped it yet either. We will update this page when that changes.

---

Sources: Anthropic, "How Claude's text watermarking works", August 2026. Anthropic Help Center, "How Claude marks AI-generated content". TechCrunch, "Anthropic says it will watermark text generated by its AI models" and "Anthropic shares more details about how Claude's new watermarks will work", August 2026.