How to Check if Claude Added a Watermark: What You Can Actually Verify

You can’t check for a Claude watermark, because Claude doesn’t add one to its text. There is no hidden character, no invisible token, no marker you can inspect in your document. Anthropic’s policy pages describe no text watermarking scheme for Claude output. What you’re seeing is a detector score, and that needs a different fix.

Why people think Claude watermarks text

The confusion is understandable. Three different things get called “watermarking,” and only one of them applies here.

Statistical watermarking is the real technique people mean. A model biases its word choices at generation time using a secret key. A detector with the same key can then test whether a passage carries that bias. It leaves nothing in the document you can open and inspect — no hidden character, no metadata field, no stray token. You cannot “check” for it in Word, in a PDF reader, or by pasting into a text editor, because there is nothing to find. It only shows up as a statistical signal, and only to a detector that holds the key.

Provenance metadata is a different thing entirely. Some image models embed signed metadata in the file itself. That is file-level, not text-level, and it is visible in the file’s properties. It has nothing to do with a paragraph you pasted into a doc.

Detector scores are the third thing, and almost always the actual source of the worry. You ran your draft through a tool, it returned a number, and someone — a teacher, an editor, a client — called that number a watermark. It isn’t. A detector score is that tool’s estimate, computed from patterns in the text. On 2026-09-28 we scored 709 passages from our own guides with the DeAIze detector at a 30% flag threshold: the mean AI score was 26%, the median 25%, and 239 of those 709 passages landed at or above the threshold. Those were unverified human-authored passages, selected by metadata only. If writing like that can score in the twenties and thirties, a mid-range reading on your draft tells you very little about who wrote it.

Can you check your own text for a Claude watermark?

No. Not with a hex editor, not with a metadata inspector, not with a “watermark checker” site.

Here is the test you can actually run. Open your document in a plain text editor. Switch on invisible characters if your editor has that option. Look for zero-width spaces, non-breaking spaces, unusual Unicode, hidden comments, and document properties. That is the complete list of things that could technically be embedded in text, and none of them is a Claude watermark. If you find something odd, it’s an artefact of copying between apps — a smart quote, a non-breaking space from a web page — not a model marker.

Any site promising to “detect the Claude watermark” is selling you a detector score with a more alarming name. That’s the uncomfortable part: the honest answer to your question is that there is nothing to check. The thing you’re actually worried about is a flag, and a flag is fixable in a way that a watermark check never was.

If a detector flagged your Claude-assisted draft, the useful next step isn’t hunting for a watermark — it’s seeing which sentences carry the signal. DeAIze scores the whole document and highlights each sentence by AI probability, so you can edit the lines that actually matter. Detection has configurable limits shown in the current interface; the rewrite trial gives you 200 words after email confirmation, no credit card. See the plans if you want the sentence-level view across a longer draft.

What Anthropic’s policy actually says

Read the policy pages rather than the forum threads. Anthropic publishes an AI detection policy and a usage policy, and those documents describe what Claude will and won’t do with your input. They do not describe a text watermarking scheme applied to Claude output. If Anthropic ships one, it will be announced — watermarking is a provenance feature, and provenance features are pointless if nobody knows they exist.

What the policy does address is how your text is handled, which is the question people often mean when they ask about watermarks. Anthropic’s stated position covers data use and retention. If your real concern is “can someone trace this paragraph back to me,” the answer runs through usage policy and account settings, not through a hidden marker in your prose.

We wrote a longer read on this in Does Claude Watermark AI Text? The Honest Answer and Your Fix, and a companion piece on removing a Claude watermark if you want the short version of why removal tools have nothing to remove.

What’s actually in your document

What you’re looking for	Does it exist in Claude text output?	How you’d check	What it tells you
Statistical watermark (biased token choices)	Reported in research on some models; not described in Anthropic’s published Claude policy	Only a detector holding the key could test it	Nothing you can act on from your side
Provenance metadata in the file	Not for pasted or exported text	File properties, metadata inspector	Whether a file was signed, not who wrote the words
Hidden characters, zero-width spaces	Sometimes, from copy-paste between apps	Plain text editor with invisibles on	Formatting artefacts, not authorship
Detector score	Yes — but it’s an estimate, not a marker	Run the detector and read its score definition	Which sentences look machine-like to that tool
Writing-style tells (flat rhythm, hedge stacking)	Yes	Read it aloud	The thing you can actually fix

That last row is where your time belongs. A detector’s output is a reading of patterns, and patterns are editable. The watermark question has no action attached to it. The style question does.

The fix: edit the sentences that carry the signal

Forget the watermark. Treat the flag as an editing brief.

1. Run a sentence-level detector. Whole-document scores hide the useful information. You want to see which specific sentences are pulling the number up, because most drafts have a handful of flat passages and a lot of ordinary prose.
2. Read those sentences aloud. Machine-made prose has a rhythm problem: uniform sentence length, stacked hedges, every paragraph the same shape. Your ear catches it faster than any tool.
3. Rewrite for specificity, not for evasion. Replace the general claim with the concrete one. Name the thing. Cut the throat-clearing clause. Add the detail only you would know.
4. Check what you changed. Rewriting can move a number, a negation, or a qualifier. Compare the revision against the original line by line and confirm every claim, quantity, citation, and condition still says what you meant.
5. Re-run the detector on the edited draft. Watch the sentence highlights, not just the headline number.

Here’s an illustration, not a measured example. Original: “It is important to consider that there are several factors which may influence the outcome of this process.” Rewritten: “Three things decide the outcome: the deadline, the reviewer, and whether the file arrived intact.” The first version is grammatically fine and says nothing. The second has a number, a list, and a point of view. Detectors tend to score the second lower because it doesn’t look like averaged prose — and more importantly, a human reader gets more from it.

That is the whole method. You are not hiding anything. You are removing the parts of your draft that were never carrying meaning in the first place.

When this is the wrong choice

Be honest with yourself about which situation you’re in.

Don’t buy an editing tool if the rule is the problem. If your course, journal, or employer prohibits AI assistance outright, a lower detector score does not make the submission compliant. Check the academic integrity policy guidance and follow the rule you’re actually under. Editing tools fix prose, not policy.

Don’t buy one if the work isn’t yours. If the ideas, data, or conclusions came from someone else, rewriting the surface doesn’t change that. DeAIze is for editing your own draft, not for passing off another person’s work.

Don’t buy one for a two-sentence caption. Sentence-level detection needs enough text to read. Short passages give you a noisy reading and not much to edit.

Don’t buy one if you just want reassurance. If you only want a number that says “you’re fine,” you’ll be back at the same anxiety next week. The point of seeing sentence-level scores is to know which lines to fix, not to collect a clean report.

Do consider it if you have a real draft of your own, you know which sections read flat, and you want a sentence-by-sentence map of where the machine-like patterns sit before you submit or publish. The rewrite trial gives you 200 words after email confirmation with no credit card, which is enough to test the sentence highlighting on a real paragraph before you spend anything.

What to do in the next ten minutes

Stop searching for a watermark checker. There is nothing in your document to find, and every tool claiming otherwise is selling you a detector score under a scarier name.

Instead: open your draft, read the flattest paragraph aloud, and rewrite one sentence so it says something only you could say. Then run it through a sentence-level detector and look at which lines still carry the signal. If you want the background on why detectors produce the readings they do, start with how AI detection works and the humanize AI text hub. Your draft isn’t marked. It’s just written in a register that reads like everyone’s.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 reference passages from guides not marked generated versus n = 12 unedited model passages, median AI score 25% and 34% respectively. Reference authorship was not independently verified; passages are not independent documents. Method: /methodology/.