Does Claude watermark text? (August 2026)

What the mark is

It is distributional, not metadata and not hidden characters. During generation the model's choice among near-equivalent next tokens is biased by a secret key, producing a statistical signature detectable across a long enough passage. The technique is the same family as Google's SynthID-Text tournament sampling.

Because the signature is carried by word choice itself, it survives copy-paste, and no character-stripping tool can touch it. Tools advertising “Claude watermark removal” by deleting zero-width characters are addressing a different thing entirely.

Scope and limits

The policy is global, driven by Article 50 of the EU AI Act, and applies across claude.ai, the API, Claude Code, Cowork and Tag. There is no opt-out and no paid tier that disables it.

Anthropic states plainly that a detected mark shows content “may have been processed by Claude” and is “not fully conclusive”, and that a missing mark proves nothing. It also notes marks may not survive heavy editing, paraphrasing, translation, format conversion or being mixed into other writing.

Detection is not available

Anthropic has not published its detection method or tooling; it says technical documentation is forthcoming. Until it does, nobody outside Anthropic can verify whether a passage carries the mark — which also means nobody can verify that a removal tool worked.

Questions

Is text I generated last week watermarked?

No. It came from a model that launched before 2 August 2026.

Will paraphrasing remove it?

In principle yes — the published research on distributional watermarks shows paraphrase attacks are the most effective removal vector, because rewriting resamples the very word choices carrying the signal. In practice nobody can confirm it against Claude specifically until Anthropic ships a detector.