Does Claude watermark its text? What it means for you
In brief
Yes. Since August 2026, text from current Claude models carries an invisible watermark in its word choice, to meet the EU AI Act. There's no opt-out, no public detector, and it carries nothing that identifies you. Here's what it is, what it isn't, and what it means at work.
Contents
Short answer: yes. Can a teacher or employer tell? Not with a public tool or a normal AI detector, but schools and universities can apply for Anthropic's detector, and light edits don't remove the mark. The safe move is to follow your school's or company's AI rules and say when you used it.
The longer version: Text written by current Claude models (Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5, plus many older ones) carries an invisible watermark. It isn't hidden characters. Claude picks between words that fit equally well in a pattern that only Anthropic's key can spot. It applies everywhere Claude runs, on every plan, and you can't turn it off. It carries no information about you, and there's no public detector: only approved groups like regulators and researchers can check for it.
Sources: Anthropic's explainer and help article.
Why is Anthropic doing this?
To comply with the EU AI Act (Article 50), which since August 2, 2026 requires AI companies serving Europe to mark AI-generated content. Anthropic applies it worldwide because, in its words, it doesn't yet have a durable way to limit it by region.
How does it work?
When Claude writes, there are often several words that would fit equally well. Normally one is picked at random. With the watermark, Anthropic's secret key decides which one, based on the words before it. Over a long passage that leaves a statistical pattern a detector with the key can find.
- Nothing is added to the text. Copy-paste, plain text and Word all keep it.
- It doesn't make the writing worse. It only changes which of several fitting words gets used.
- No extra cost or speed hit.
It's based on Google DeepMind's SynthID-Text approach.
Which models and products?
- Models: every current model, and older ones going back to Opus 4.5, Sonnet 4.5 and Haiku 4.5 (added through November 2026).
- Products: the Claude apps, Claude Code, Cowork, Claude Tag and the API.
- Clouds: the same watermark applies on AWS, Google Cloud and Microsoft Foundry.
Developers: on some older models (Opus 4.6 and 4.5, Sonnet 4.6 and 4.5, Haiku 4.5), setting temperature, top_p or top_k to a non-default value can weaken the mark. The current models don't have that note.
Who can detect it?
Not the public. The detection tool is in private preview for groups the EU law has in mind: regulators, law enforcement, media and fact-checkers, researchers, schools and universities, EU civil society groups and some companies with their own legal duties. They apply through a form.
Third-party "AI detectors" don't have the key, so they can't read this watermark. Anthropic's free Content Checker is public, but it only reads the labels on image files. It can't check text.
Even a positive result isn't proof. Anthropic calls a detected mark "a signal" that Claude processed the text, "not fully conclusive".
Does editing remove it?
It depends how much you change:
- Light edits usually leave it in place.
- A full rewrite in your own words removes it.
- Short passages, code and dense factual text carry a weak mark that is hard to detect.
- Translations by Claude are marked, since Claude chose every word.
What if I wrote it and Claude only edited it?
Only the words Claude chose carry the mark. A light proofread usually changes too few words to register. If Claude rewrote your draft, the rewritten parts can carry it, even though the ideas were yours.
What about images and files?
Images and files Claude makes (like PNG and JPEG) get a different kind of label: signed metadata (C2PA Content Credentials) attached to the file. Re-saving or taking a screenshot removes it.
What it means at work
- Nothing changes in how you use Claude. Your rights to the output are the same.
- It isn't tracking. The mark can't be traced to a person, organization or chat.
- Be open about AI help where it matters, such as client work, school, publishing or regulated content. That was good practice before the watermark, and now some readers may be able to check.
- Admins: there's no setting for it on any plan. Add one line to your AI usage policy saying when people should disclose AI help.
The backlash
When it was announced in August, some subscribers objected that their own reworked writing could be flagged, and a few said they'd cancel. Anthropic's answer was that the mark carries no identifying information and that light edits by Claude rarely register.
Quick answers
Does Claude watermark its text? Yes. Since August 2026, text from current Claude models carries an invisible statistical watermark in its word choice, to comply with the EU AI Act. It applies on every plan and platform worldwide.
Can I turn off the Claude watermark? No. There is no opt-out on any plan, in the API or on cloud platforms.
Can my teacher or employer detect Claude text? Not with a public tool. Detection is in private preview for approved groups such as regulators, researchers and educational organizations, and third-party AI detectors can't read Anthropic's watermark.
Does the Claude watermark identify me? No. Anthropic says it carries no identifying information and can't be traced to a person, organization or chat.
Does editing Claude's text remove the watermark? Light edits usually don't. A full rewrite in your own words does. Short passages and code carry a weak mark that is hard to detect.