The report
What the Claude watermark detector reports
This is the inspection step before text goes somewhere that treats characters literally. Press Scan text and the paste comes back marked up in place, with two independent counters underneath: invisible Unicode characters found, from a fixed list of 14, and writing patterns matched, from a list of 17. Every marked item carries its own reason on hover.
Everything runs in the page. No upload, no account, no character cap, no scan limit. The scan itself is read-only, so your text is unchanged unless you choose to press Clean it afterwards.
Scope
The one thing it cannot check
Settle this before you rely on it for anything. Claude's text watermark is not detectable by this tool or by any public tool. Anthropic's help centre describes an imperceptible watermark placed into the text itself, which travels with the text through copy and paste, and states that watermark detection is currently in private preview, limited to eligible organisations under EU law.
Checking it requires the key that encodes it. No third party holds that key. Anthropic's own free checker that anyone can use is a file checker: it reports whether a file carries a Claude-issued Content Credential under the C2PA standard, which is a different question from anything a text box can answer.
How to
How to run the Claude watermark detector in three steps
Scan the text in the state it arrived. Any retyping or reformatting you do first will quietly remove the evidence you were looking for.
- Step 1
- Paste the text. No cap applies, so a whole file works. Try a sample loads a dirty example if you want to see the report populated.
- Step 2
- Press Scan text. Read the two counters separately and hover the marked items for the reason behind each one.
- Step 3
- Decide. Clean it strips the characters and leaves the words alone. If the count was zero, there is nothing to do.
Triage
What to do with each kind of finding
The two counters call for different responses, and treating them as one list is where people go wrong.
- Zero-width characters
- Remove them. They serve no purpose in pasted prose and they break exact matching everywhere downstream.
- Space look-alikes
- Also remove, but note that these are replaced with an ordinary space rather than deleted, so word boundaries survive.
- Byte-order mark mid-document
- Remove. It is only meaningful at the very start of a file and causes parser errors anywhere else.
- Direction controls
- Worth looking at rather than removing blindly. If your text is genuinely bidirectional, some of these belong there.
- Flagged phrases
- Do not treat as errors. They mark generic writing, so the response is a rewrite, and cleaning will not touch them.
Workflow
When the scan is worth the step
It is local and unmetered, so the cost is a couple of seconds. Still, it is only useful at certain points.
- Text entering a repository
- A stray character makes git report a change nobody can see, and it survives review for the same reason.
- Text entering a CMS
- Slugs, tags and meta fields match literally, so one invisible character makes a URL resolve to nothing.
- Text entering a sheet or database
- Lookups fail against values that are visually identical. This is the failure that takes longest to find.
- Text you received from someone else
- You do not know the copy path it took. A scan tells you before you inherit the problem.
- Not before a human reads it
- If the text is only going to be read, none of these characters matter. Skip it.
Reading it
Reading a zero result correctly
A zero means no invisible characters were found in that paste. It is a real result and often the right one, because plain text copy paths tend to be clean.
What it does not mean is that the text carries no Claude watermark, because the watermark is never examined here. Anthropic makes the equivalent point about their own detection: the absence of a mark does not mean content was not AI-generated or processed, and a detected mark is a signal rather than a conclusion. A green screen only answers the question the tool asked.
Limits
What the Claude watermark detector will not tell you
It will not tell you who wrote the text. Invisible characters come from the copy path, not the author, and the phrase patterns occur in human writing constantly. There is no verdict, no probability and no score anywhere in the output.
It also inspects a narrow surface on purpose. No punctuation analysis, so quotes, en dashes and em dashes are not examined or reported. No HTML inspection. No file input, no batch mode, no API. And no prediction about what any AI detector will say about your writing.
In practice
Who puts a scan in the pipeline
Engineers moving assistant output into code and configuration. Content teams accepting drafts from contributors whose copy route they do not control. Analysts pasting generated lists into spreadsheets. Editors who want to see what a document is carrying before they touch it.
The habit that works is separating inspection from action. Scan, read the reasons, then decide whether to clean. Running a cleaner first and asking questions later removes the information you would have used to decide.