Check
Claude Watermark Test
Claude watermark test tools all measure something, but for text none of them measure the watermark.
Knowing what each test actually reads is the difference between useful information and false reassurance, and the gap between those two is where people get into trouble.
Run the metadata test on files. For text, no valid test is available to anyone yet.
Runs in your browser. No signup, no upload, nothing stored.
Tests you can run and trust
These read data that is genuinely present in the file and report it accurately:
- C2PA manifest read on .png, .jpg or .svg
- EXIF and XMP metadata dump on any image
- Invisible-character scan on text
What the character scan proves
It proves whether zero-width and formatting characters are present in your text. That is genuinely worth knowing, because they break search, confuse diffs and occasionally break code.
It proves nothing at all about the Claude watermark, which does not use them. Two useful facts that are frequently reported as one.
Tests that look valid and are not
AI-writing classifiers return a confident-looking percentage derived from stylistic features — sentence length variance, vocabulary distribution, punctuation habits.
They were unreliable before watermarking existed, and watermarking did not improve them. A high score is a statement about how the text reads, not about where it came from.
A test you can run on yourself
Take a passage you wrote entirely by hand, with no assistance of any kind, and run it through the classifiers you were about to trust.
People routinely find their own unaided writing scored as machine-generated, particularly if they write cleanly or in a second language. That result tells you exactly how much weight the tool deserves.
Testing before you publish
The useful pre-publication test is not a scanner. It is the provenance question: which model produced this, when, and how much has changed since.
That reasoning is available today and gives you something to act on, where a scanner gives you a number with no defined meaning.
When a real test arrives
Anthropic has committed to supporting detection by users and third parties, without a date.
When that ships, it will report a likelihood rather than a verdict, and it will apply retroactively to everything published since 2 August 2026. Test results you get today do not change that.
Test the file instead
If your question is about an image rather than a passage of text, you are in luck: the C2PA manifest is a real, readable record and the test is definitive.
That asymmetry is worth exploiting. Where a file exists, check the file — it answers in seconds what no amount of text analysis currently can.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.