Does Claude watermark its text? What it means for your work
Anthropic's Claude text watermark hides a pattern in word choices to meet the EU AI Act. What it is, who can detect it, and what it means for you.
Some Claude models do. Anthropic's release notes say text from Claude Fable 5.1 and Claude Mythos 5.1 carries a text watermark, a hidden pattern in the words the model picks. It does not name you, and Anthropic reports no effect on quality. Detection is limited to eligible organizations.
This site is independent and not affiliated with Anthropic. We read Anthropic's post and release notes on 2026-10-21. We did not run a detector, because we found no public one, and we did not test how the watermark behaves on real text.
What the watermark is
Anthropic published an explainer on August 14, 2026. It says Claude's watermark is "a version of the SynthID-Text approach published by Google DeepMind". The idea starts from a plain fact: a model picks one word at a time, and often several words would fit equally well.
Anthropic says that in those cases "the choice is settled by a random number". A watermark steers that choice, and anyone holding the key can later "check the sequence of words" against what the model would have picked. Anthropic's watermark explainer has the full description.
Which models carry it
The explainer says "Future Claude models will generate text that contains a watermark". The Claude Platform release notes are more specific, in an entry dated September 1, 2026:
Text generated by Claude Fable 5.1 and Claude Mythos 5.1 carries Anthropic's text watermark
We found no statement for Claude Opus 5.5, Sonnet 5.5 or Haiku 5.5, so we do not claim either way. If it matters for your work, check the release notes again. Our model comparison lists the current lineup.
What it changes for you
Very little, day to day. Anthropic says "no impact of watermarking on the content, level of creativity, or readability" of Claude's text. That is Anthropic's own testing, and we could not check it.
On privacy, Anthropic says the watermark does not identify "anything to do with individual users". It marks that Claude wrote the words. It does not record your name, your account or your prompt.
Anthropic also says code carries less of it: code "has generally less watermarking than some other forms of text". Code often has to be exact, so there are fewer free word choices to nudge.
The Claude moves worth stealing, by email
Workflows tested on real jobs, sent only when they are worth it. Free.
Edited and proofread text
This is the case most readers ask about. Anthropic's answer for text you wrote yourself and asked Claude to proofread is that there is "very little (if anything) for the watermark to attach to". Most of the words are yours.
The reverse follows by the same logic, though Anthropic's post does not spell it out for every case. A long passage Claude wrote and you pasted unchanged keeps the most pattern. A passage you rewrote in your own words keeps less. Treat that as our reading, not a measured result.
Who can detect it
Anthropic describes a detection tool in "private preview", open to eligible organizations "as required under EU law". The post names regulators, law enforcement, media, fact-checkers and independent researchers among them. We found no public checker where you can paste text and get a verdict.
For images and other files, Anthropic uses a different mark. The explainer says Claude attaches a "cryptographically signed note" to supported files, saying they were made with Claude. The release notes call these C2PA Content Credentials. Our guide on whether Claude can generate images covers what Claude produces as files.
Why Anthropic does this
Anthropic gives one reason: it is acting "to comply with the EU AI Act". The explainer presents the watermark as a legal compliance step, not as a feature you asked for.
What to do with this
You do not need to change how you use Claude. Two habits help.
- Check your rules. A watermark is a technical mark. Your school, client or employer sets the actual policy on AI help. If you are unsure, ask before you submit.
- Own the final text. Read it, fix names and numbers, and change what does not sound like you. Our what is Claude explainer covers why that check matters for any draft.
Limits of this brief
- We rely on Anthropic's own account. We could not run a detector or compare watermarked and plain text.
- The explainer is dated August 14, 2026 and the release note September 1. Both may have moved on since.
- The post does not say how much editing removes the pattern, so we give no number.
- We do not know which models beyond Fable 5.1 and Mythos 5.1 carry it today.
Does Claude watermark text?
Some models do. Anthropic's release notes say text from Claude Fable 5.1 and Claude Mythos 5.1 carries its text watermark, a hidden pattern in word choices. We found no statement on Opus 5.5, Sonnet 5.5 or Haiku 5.5, so check the release notes.
Can I tell if text came from Claude?
Not on your own. Anthropic's detection tool is in private preview for eligible organizations such as regulators, law enforcement and researchers. We found no public checker where you can paste text and get an answer.
Does the Claude watermark identify me?
No, according to Anthropic. It says the watermark does not identify "anything to do with individual users". It marks that Claude wrote the words, not who asked for them. We cannot verify that claim independently.
Does editing Claude's text remove the watermark?
For your own text proofread by Claude, Anthropic says there is "very little (if anything) for the watermark to attach to". For a passage Claude wrote, the post gives no threshold, so do not assume a set amount of editing makes it undetectable.
Sources (checked 2026-10-21)
- How Claude's text watermark works: Anthropic's explainer, August 14, 2026.
- Claude Platform release notes: the September 1, 2026 entry naming Fable 5.1 and Mythos 5.1.