Does Claude watermark text?
Yes. Anthropic began embedding invisible watermarks in Claude’s text output on 2 August 2026.
The driver was the EU AI Act, specifically the Article 50(2) Code of Practice on Transparency of AI-Generated Content, which requires generative AI providers to mark output so it can be identified as machine made. Anthropic signed it.
The rollout was not limited to the EU. It applies worldwide, to everyone using Claude, regardless of where they are.
Which models and surfaces carry it?
The boundary is the launch date of the model, not the date you used it.
| Marked | |
|---|---|
| Claude models launched on or after 2 Aug 2026 | Yes |
| Claude models launched before that date | Not yet, transition period with no announced end |
| Claude apps | Yes, when the model is a marked one |
| Claude Platform / API | Yes |
| Claude Code | Yes |
| Claude Cowork, Claude Tag | Yes |
| AWS, Google Cloud, Microsoft Foundry deployments | Yes |
This matters more than it sounds. Output from an older model may carry nothing at all, so absence of a mark tells you very little.
What kind of watermark is it?
A statistical one. At each generation step the model’s choice among statistically near equivalent next tokens is nudged according to a secret key. The signal lives in the pattern of word choices across a passage.
Anthropic’s own framing is that the mark does not change the meaning, quality or readability of the response.
It is not zero width characters, and it is not Unicode variation selectors. That kind of marking would be visible in a hex editor and removable in one regular expression, which would make it useless as provenance. Any page telling you Claude’s watermark is invisible characters has the mechanism wrong.
What does it survive?
- Copy and paste: survives.
- Light editing: survives.
- Deleting invisible characters: unaffected, because it is not made of them.
- Heavy paraphrase or a real rewrite: degrades, and may drop below the detection threshold.
- Very short passages: never reliably detectable, because there is not enough signal.
- Code formatters and linters: degrades, which matters most for generated code.
Can anyone actually check?
Not yet. Detection tooling was described as forthcoming, with no public API, no access model and no published technical documentation. Right now nobody outside Anthropic can test a passage.
Sources also disagree on how broadly the marking is live today. Some describe it as deployed across new models; others say no publicly checkable watermark has shipped. Treat confident claims in either direction with suspicion, including confident claims that it is easy to remove.
What a mark would prove, if you could check
Less than people assume.
- A detected mark does not prove AI authorship. Claude may have edited, summarised or proofread text a person wrote.
- No mark does not prove human authorship. The model may predate the rollout, or the passage may be too short.
Anyone planning to treat detection as evidence in a classroom, a newsroom or an HR process should sit with those two limits first.
Who this is not for
If you came here looking for the trick that strips the mark, there isn’t one, and the pages selling it are describing invisible characters instead. What is real and worth doing is cleaning the invisible characters that break your build, which is a different job with a different reason.
For the full technical breakdown, see what Claude’s text watermark actually is.