Claude watermark detector
What a Claude watermark is, what can be tested today, and what cannot be — stated plainly, because the difference matters more than the score.
FreeAIText does not currently have access to Anthropic's private watermark key or detector. Today's score is a local statistical estimate, not an Anthropic verdict.
The short version
Anthropic has announced model-level text watermarking for Claude. The watermark is statistical: it biases word and token selection using a keyed source of randomness, so a party holding the key can test whether a sequence matches the choices Claude would have made. Anthropic states that nothing is added to the text and there are no hidden characters.
Anthropic has also said it will offer a watermark detection API, and that the details are still being worked out. Until that ships, no third party can verify Claude's watermark — the key is Anthropic's.
| Claude watermark | Announced — statistical, word/token selection |
|---|---|
| Hidden characters | None, per Anthropic |
| Public detection API | Not publicly available |
| What FreeAIText runs today | A local research estimator. Not Anthropic's detector. |
| Last checked | 18 August 2026 |
What the local estimator measures
The tool on this site reproduces the green-list z-test from the published watermarking literature. It splits your text into segments, decides green-list membership for each token from a hash of the preceding one, and reports how far the resulting count sits from chance.
Because the key is not any vendor's, the result is a measure of general statistical watermark pressure. It is a research instrument. It is not Claude's detector, and a high score is not an Anthropic verdict. Highlighted segments are the ones contributing most to that local estimate — not token attribution from Anthropic's detector.
Reading a watermark result honestly
A positive result indicates possible involvement, not authorship. Text where a person wrote the substance and a model rephrased it carries the same statistical signature as text pasted verbatim.
A negative result does not prove human authorship. Unwatermarked models, text generated before a watermark rolled out, short samples and heavily edited documents all return nothing.
Some samples cannot be judged. Short text lacks the statistical power for a claim in either direction; heavy editing and translation dilute or replace the token sequence. Treat those results as inconclusive.
The longer version, with the reasoning: what a Claude watermark does and does not prove.
What this is not
- Not an AI-writing classifier. Those guess from style and misfire badly on non-native English. See AI detector vs watermark detector.
- Not a Unicode cleaner substitute. Invisible-character stripping is a different layer, and Claude's watermark does not live there — why Unicode cleaners do not remove Claude watermarks.
- Not a guarantee of anything. Nobody can promise a result is permanently or universally undetectable, and we will not imply it.
- Not an Anthropic product. FreeAIText is an independent service and is not affiliated with, endorsed by, or sponsored by Anthropic.
When the API arrives
The integration is already built to the boundary: a provider interface, a normalisation layer so the interface does not depend on Anthropic's future response schema, and an explicit feature flag that has to be switched on deliberately. Nothing here will describe a score as API-verified until a real API response has been received.
Progress and the current date-stamped status: Anthropic watermark detection API status.