Does Anthropic watermark Claude's text? Yes — since August 2026
Anthropic now embeds an invisible watermark in Claude's text output. What it is, what it proves, and what Anthropic itself says degrades it.
Anthropic's watermark · 4 min read
Yes. As of 11 August 2026, Anthropic has confirmed that Claude embeds an invisible watermark directly into generated text.
This is new. Until this month no major provider watermarked text output except Google, and anyone telling you otherwise about Claude was working from stale information.
What Anthropic has actually shipped
From Anthropic's own documentation and its announcement:
- Every Claude model launched on or after 2 August 2026 marks generated text at launch. Support for models released before that date is described as "in progress".
- The mark is applied at the model level, so it is present regardless of which Claude product or surface the text came from — app, API, or anything built on top.
- It is imperceptible: woven into the text itself, not visible, and does not change meaning or readability.
- It is applied globally, not only to users in the EU.
- For supported image files (
.svg,.png,.jpg) Anthropic uses the C2PA standard — signed provenance metadata, a different mechanism from the text mark.
The immediate driver is EU AI Act Article 50, which became enforceable on 2 August 2026 and requires providers of generative AI systems to embed machine-readable marks in their outputs, text included.
What the mark proves — and doesn't
This is the part most coverage gets wrong. Anthropic is explicit that a detected mark indicates content may have been processed by Claude. It does not establish that:
- Claude authored the original ideas
- the content is unchanged since
- a human did not write the substance and use Claude to edit it
Processing, not authorship. Text you wrote yourself and asked Claude to tidy up carries the same mark as text Claude wrote from scratch.
How durable is it
Anthropic's own documentation sets the boundaries, and they are worth quoting accurately rather than paraphrasing.
It does survive: copying and pasting the text elsewhere. Anthropic says the mark "will travel with the text", and "may persist through some editing".
It may not survive:
- heavy editing
- paraphrasing
- translation
- mixing with other content
- very short passages, which lack enough text for reliable detection
Note "may" in both directions. Anthropic is not promising the mark survives light edits, and not promising it fails under heavy ones. Anyone who states either as certainty is guessing.
The detection mechanism is not public yet
Anthropic says it will "share details on detection mechanisms in forthcoming technical documentation". Right now that documentation does not exist.
The practical consequence is significant and rarely stated: nobody outside Anthropic can currently check whether a given piece of text carries the mark. Not you, not us, not a university, not a publisher.
Which means any product claiming to verify it has removed Anthropic's watermark is claiming something it cannot demonstrate. Including this one. We will say plainly what we do and do not know:
- Verifiable: hidden characters, exotic spacing and typographic punctuation. We count them exactly and you can re-scan to confirm zero.
- Not verifiable by anyone today: whether a statistical text watermark is present or absent.
When Anthropic publishes a detector, that changes and we will test against it.
What this means if you use Claude
If the mark matters to you, the honest picture is:
- Character cleanup does not touch it. Stripping zero-width characters and em dashes is worth doing for other reasons, but a statistical watermark woven into word choice is not a character you can delete.
- Paraphrasing is the mechanism Anthropic itself names. A watermark that lives in which words were selected degrades when the words change. That is why "paraphrasing" appears on Anthropic's own list.
- Depth matters. "Heavy editing" is on the list; light editing is not. A one-word-per-paragraph touch-up is not what Anthropic is describing.
- Short text is unreliable anyway, by Anthropic's own admission.
Why a watermark behaves this way
Statistical text watermarking works by biasing which token the model picks among several near-equivalent options, so the accumulated pattern is detectable by a key holder. Google's SynthID-Text, published in Nature in 2024, works this way and is the best-documented example.
The signal therefore lives in the specific words chosen. Change the words and you change the signal. This is not a flaw in one implementation — it is inherent to the approach, and it is why OpenAI held its own text watermark back rather than shipping it.
Scan your text to see the character-level markers exactly. On the statistical layer we will tell you what is known and not pretend to more.
Sources: Anthropic Help Center: how Claude marks AI-generated content · TechCrunch, 11 Aug 2026 · Fortune · Forbes
Related: EU AI Act Article 50 explained · what SynthID does