Claude watermark
Anthropic marks Claude’s text with a version of SynthID-Text, biasing which word the model picks next. Nothing is inserted into the text, so no character scanner can find it.
What the evidence says
Anthropic marks the text of Claude models launched from 2 August 2026, the date the EU AI Act transparency rules took effect. Marking applies worldwide rather than only in Europe, and covers the API, Claude, Claude Code and the cloud platforms that resell the models. Models released before that date are in a transition period and are not marked yet, so the answer for a given passage depends on which model produced it.
The method is a version of SynthID-Text, the scheme Google DeepMind published and open-sourced — the same family of watermark that marks Gemini. Instead of using an arbitrary random number generator to pick the next word, the model uses a key and the few words that come before to settle the choice, so the sequence of choices itself carries a pattern that a holder of the key can measure.
That mechanism decides what any tool can do about it. Nothing is added to the text and there are no hidden characters, so scanning a document for zero-width spaces, byte-order marks or unusual spacing tells you nothing about whether Claude wrote it. Anthropic also signs C2PA provenance metadata on generated image files, which is a separate mechanism and does not apply to prose.
The stated limits are specific. Detection works poorly on short samples, where there are fewer word choices to measure. It is sparser on factual passages, and where an exact output is required — code, or a term that would be wrong if it were swapped — the watermark is not applied at all. Light editing probably will not remove it; a complete rewrite where every word is replaced will. When Claude edits someone else’s writing, most of the words are the person’s and there is very little for the mark to attach to.
Detection is not open to the public. The detection API is in private preview for organisations eligible under EU law, such as regulators, law enforcement, media, fact-checkers, independent researchers, educational institutions and EU civil society groups. Anthropic documents no way to switch marking off. It also cautions that a detected mark is a signal rather than proof, and that the absence of one does not mean a text was not AI-generated.
A common misreading
Sites that sell “Claude watermark removers” strip invisible Unicode characters. Anthropic’s watermark is not made of characters, so removing them does not touch it. Those characters may still be worth cleaning up, but not for that reason.
Sources
- How Claude’s text watermarking worksAnthropic · · Provider documentation
- How Claude marks AI-generated contentAnthropic Help Center · · Provider documentation
- Anthropic shares more details about how Claude’s new watermarks will workTechCrunch · · Independent report
Frequently asked questions
Can I detect Claude’s watermark myself?
No. Detection needs the key, and the detection API is in private preview for organisations eligible under EU law. No public tool, including ours, can read this signal from text alone.
Do invisible characters in my text mean Claude wrote it?
No. Invisible characters come from editors, clipboards, PDFs and web pages as readily as from any model, and Claude’s watermark does not use them. Finding one points to a text-handling step, not to an author.
Is every Claude response watermarked?
No. Marking covers models launched from August 2026, and it is not applied where an exact output is required, such as code. Older models are in a transition period and are not marked yet.
Related tools
Working on your own draft?
GenPolish rewrites text you already have, showing every change before you accept it. It does not add or remove provenance signals, and it makes no claim about detectors.