Is your Claude text watermarked?
Short answer: not yet — no shipped Claude model carries Anthropic’s new mark. But your text is very likely still identifiable right now, for a completely different reason. Paste it in and see exactly what a detector would find.
What Anthropic actually announced
On 11 August 2026 Anthropic confirmed that Claude will embed a machine-readable watermark directly into generated text, under Article 50 of the EU AI Act. It applies worldwide, across claude.ai, the API, Claude Code, Cowork and Tag, with no opt-out and no paid tier to disable it.
The mark is not metadata and not invisible characters. It is distributional: the model’s choice among near-equivalent next words is nudged by a secret key, so the watermark is the word choice. That is why it survives copy-paste — and why no amount of character-stripping touches it.
Anthropic’s own documentation is unusually candid about the limits: a detected mark means content “may have been processed by Claude” and is “not fully conclusive”, while an absent mark proves nothing at all. Marks may not survive heavy editing, paraphrasing, translation, or being mixed into other writing.
Questions
- Does Claude watermark its text?
- Anthropic announced that Claude models launched on or after 2 August 2026 embed an imperceptible statistical watermark. But no shipped model qualifies yet — Opus 5, Fable 5 and Opus 4.8 all launched before that date. Anthropic has said it is working to add marking to earlier models. So as of today, text from production Claude models does not carry the mark.
- Can anyone detect a Claude watermark right now?
- No. Anthropic has not published its detection method or tooling — it says technical documentation is forthcoming. Any tool claiming to detect Anthropic's statistical watermark today is guessing. We say so rather than pretending otherwise.
- So what is actually detectable in my text?
- Copy-paste artifacts. When you paste straight out of a chat UI into a rich-text editor, provider-specific HTML classes and attributes come with it — Claude stamps class names containing 'claude'. Zero-width characters, non-breaking spaces and typographic punctuation also survive. These are deterministic and easy to grep for, which is exactly how the AO3 Claude detector works.
- How does the AO3 Claude detector work?
- It is a single CSS rule that matches paragraphs whose class name contains 'claude' inside a work's body. It has effectively no false positives, because that class can only be present if the text was pasted directly from Claude without passing through another editor. Cleaning the HTML removes it.
- Is the checker free?
- Yes, and unlimited, with no account. It runs entirely in your browser, so your text is never uploaded. Only the optional rewrite engine costs credits, because only it costs us money to run.
- Will you remove the watermark when it ships?
- That is what the rewrite engine is for. A statistical watermark lives in word choice, so the only thing that removes it is resampling the text with a model that does not itself watermark. Our engine already runs that stage as a no-op, and it activates automatically once Anthropic ships a marked model and a detector we can verify against.
Unmarked is independent and not affiliated with Anthropic.