How to check if text has Claude's watermark
What you can detect yourself, what no public tool can detect, and why every 'Claude watermark checker' claiming a percentage is guessing.
Removing the watermark · 2 min read
If you want to check whether text carries Claude's watermark, the honest answer splits into something you can do precisely and something nobody can do at all.
What you can check exactly
Character-level markers. These are real codepoints in the text, so detection is a lookup rather than a judgement:
| Category | What to look for |
|---|---|
| Zero-width | U+200B, U+200C, U+200D, U+2060, U+FEFF, U+00AD |
| Tag characters | The U+E0000 block — no legitimate use in prose |
| Variation selectors | U+FE00–FE0F, U+E0100–E01EF |
| Exotic spacing | U+00A0, U+2000–200A, U+202F, U+3000 |
| Typographic punctuation | Em dash, en dash, curly quotes, U+2026 |
A scan returns counts, not probabilities. If it says three zero-width characters, there are exactly three. Run one free — it executes in your browser, so the text never leaves your machine.
What nobody can check
The statistical watermark itself.
Anthropic has not published a detection mechanism. Its documentation says technical details are forthcoming. Until then there is no public detector, which means:
- No website can tell you whether text carries Claude's mark.
- No browser extension can.
- We cannot.
If a "Claude watermark checker" gives you a percentage, it is either running a generic AI-detector heuristic — perplexity and burstiness, which is guessing about style, not reading a watermark — or making the number up. Those are the only two options, because the information required is not publicly available.
How to tell the difference
Ask what the tool claims to measure.
"We found 12 zero-width characters" — checkable. Ask it which codepoints and where; a real scanner can tell you.
"This text is 87% likely to be Claude-generated" — a style guess. It might correlate with reality, and it produces false positives on non-native English writers, technical writing and heavily edited prose. It is not reading a watermark.
"Watermark detected" — not currently possible for Claude. Nobody has the detector.
What the character scan tells you in practice
A scan of Claude output almost always shows em dashes, curly quotes and the U+2026 ellipsis. That tells you the text probably came from a model — a strong stylistic signal, but not the watermark, and not proof.
Zero-width or tag characters are more interesting. Those have no business in ordinary prose, and an even distribution through a document does not look like a formatting artefact.
What to do with a clean result
A zero count means no hidden characters and no obvious typographic tells. It does not mean no watermark. A statistical mark is invisible to character analysis by design.
Being precise about that boundary is more useful than a reassuring number, which is why we report counts and refuse to report a watermark probability.
Related: how to remove Claude's watermark · auditing a document · how AI detectors work