Claude watermark detector
What the Claude watermark detector checks
Paste text, press Scan text, and the report comes back marked up. Two counters sit under it. One is the number of invisible Unicode characters found, drawn from a fixed list of 14. The other is the number of writing patterns matched, drawn from a list of 17. Hover any marked item and it tells you what that thing is and why it matters.
It runs in your browser. Nothing is uploaded, there is no account, no cap on length and no limit on how often you scan. If you want to remove what it found, the same widget has a Clean it button, but the scan itself changes nothing.
Read this first
What no page can detect right now
Claude's text watermark is not something this or any other public tool can find. Anthropic's help centre describes it as an imperceptible watermark placed directly into the text, which travels with the text when it is copied and pasted, and says watermark detection is currently in private preview rather than publicly available. Access is limited to eligible organisations under EU law: regulators, law enforcement, media, fact checkers, researchers, educational organisations and EU civil society groups.
Detection needs the key that encodes the mark. We do not have it. Neither does any of the free checkers competing for this search term, whatever their result screen looks like. Anthropic's separate free Content Checker is for files rather than pasted text: it tells you whether a file carries a Claude-issued Content Credential under the C2PA standard.
So this tool detects invisible characters. It is a different thing, it is a real thing, and it is worth knowing what is in your text.
How to
How to use the Claude watermark detector in three steps
Scan the text as it arrived, before you edit it. Retyping a sentence removes whatever was in it, so an edit made first hides the thing you were trying to find.
- Step 1
- Paste the text. There is no length limit. Try a sample loads a deliberately dirty example if you want to see a populated report.
- Step 2
- Press Scan text. Findings are marked in place, and the two counters below give hidden Unicode characters and flagged phrases separately.
- Step 3
- Hover each finding for its reason and decide item by item. Press Clean it only if you want the characters taken out.
The character list
The 14 characters it looks for
The list is fixed. It covers the characters that actually turn up in pasted text and cause trouble, not every unusual codepoint in Unicode.
- Zero-width space and friends
- Zero-width space, zero-width non-joiner, zero-width joiner and word joiner. Each takes a character slot and renders nothing, so the string you have is not the string you see.
- Byte-order mark
- Fine at the very start of a file, a problem anywhere else. Mid-document it upsets parsers, feeds and validators.
- Soft hyphen
- Invisible until the line wraps, when a hyphen appears in a word that never had one.
- Three space look-alikes
- No-break space, narrow no-break space and hair space. They read as ordinary spaces and are not.
- Five direction controls
- Two embeddings, a pop, and the two overrides. An override can make displayed text read differently from what is stored, which matters in file names and links.
- What it deliberately skips
- A zero-width joiner between two emoji is doing its job, so it is not reported. Only stray ones are.
The second count
The 17 phrase patterns, and what they are not
The second counter is a different kind of finding. It matches 17 writing patterns common in assistant prose: stock openers, vague importance, hype adjectives, negation pivots, tour-guide phrasing, overused verbs. Overlapping matches are merged, so the number is spans of text rather than raw hits.
These are prompts for an edit, not evidence of anything. Human writers use every one of them. A high phrase count on a paragraph means the paragraph reads generically, which is a good reason to rewrite it and a bad reason to conclude anything about who typed it.
The trap
A clean scan does not mean clean text
This is the thing to understand about every checker on this search term. A zero result here means no invisible characters were found. It says nothing whatsoever about Claude's watermark, because the watermark was never examined.
Anthropic makes the same point from the other side 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. Treat a green result on any tool as an answer to the question that tool actually asked.
The blunt question
Does the Claude watermark detector prove Claude wrote it?
No. An invisible character tells you something about the copy path, not about the author. A zero-width space can arrive from a web page, a word processor or a note app, and plenty of Claude output has none in it at all.
The phrase findings prove even less. We do not give a verdict, a probability or a score, we make no prediction about any AI detector's result, and we have no affiliation with Anthropic.
In practice
Who scans text, and when
Developers checking a paste before it enters a repository. Editors who want to know what a submitted document is carrying before they work on it. Anyone whose find and replace, URL slug or spreadsheet lookup has started failing on text that looks completely normal.
Scan first, decide second. The report exists so you can see what is there and choose, rather than running a cleaner over text that had nothing wrong with it.