Claude now watermarks
everything it writes.
It marks that a model touched the text. It does not say who wrote it. That gap is about to cost honest writers.
What actually shipped.
Anthropic announced on August 11, 2026 that Claude embeds an invisible mark in the text it generates.
What it is
A statistical watermark applied while the text is being written. The model is nudged between wordings that are near enough to equivalent that the meaning does not change, in a pattern a matching detector can recognize later.
Where it applies
Everywhere Claude is offered, not only in the EU. Because the mark is applied by the model itself, it covers the apps, the API, the coding tools, and Claude served through other cloud platforms.
Which models
Models released on or after August 2, 2026. Earlier models are being updated, so an unmarked passage may simply have come from an older one.
Why now
The EU AI Act's transparency rules became binding on August 2, 2026, and require generated content to be machine readable as generated. The rule is European. The implementation is global.
What survives
Copying and pasting keeps the mark, because it lives in the word choices rather than in hidden characters or file metadata. Light editing usually keeps it too.
What breaks it
Heavy rewriting, translation, and summarizing all weaken it. Short passages never carry enough signal to read at all, so a few sentences cannot be checked either way.
A mark is not a verdict.
The watermark answers a narrow question accurately. The trouble starts when a reader treats the answer as covering a much wider one.
What a mark tells you
- Text passed through a Claude model that supports marking.
- It passed through for a long enough stretch to leave a readable signal.
- Something in the chain between the blank page and the final draft involved that model.
What it cannot tell you
- Who wrote the argument, did the research, or produced the first draft.
- Whether the model wrote the passage or only tightened wording you had already written.
- That absent marks mean human writing. Older models, short passages, rewriting, and translation all come back clean.
If your own writing carries a mark.
The people hit hardest by this are the ones who used Claude to proofread work that was entirely theirs. If that is you, the mark is real and the accusation behind it is still wrong.
Separate the two claims before you answer
A mark is a claim about processing. An accusation is a claim about authorship. They are not the same claim, and the mark cannot support the second one. Say that plainly and early, because most of these conversations go wrong when both sides treat one number as settling everything.
Show the record, not the score
The brief, dated notes, an outline with abandoned ideas still in it, version history, and the edits you made after feedback. That record answers the question the watermark cannot, and it is far more persuasive than arguing about a detector.
Be specific about what the tool did
There is a real difference between drafting a section and fixing comma splices in one you wrote. If you used Claude to proofread, say so, and point to the draft that existed before you did. Precision reads as confidence.
Do not rewrite good prose to defeat a check
Adding errors and odd phrasing to look human makes the writing worse and still proves nothing about who wrote it. If a flagged passage is genuinely stale, fix it on editorial grounds. Otherwise leave it alone.
Who marks text, and who does not.
Coverage is partial, and the gaps are large enough that a watermark check alone will never be a complete answer.
Claude
Marks text
Live since August 2, 2026 on models released from that date, worldwide, across every Claude product.
ChatGPT
No text mark
OpenAI built text watermarking and chose not to turn it on, citing the risk of penalizing people who use it as a writing aid. Images carry provenance data.
Gemini
Marks text
Google's text watermarking has existed for a while, but reading it requires Google's own key, so there is no public way to check a passage.
Open models
No mark
Anything self-hosted from open weights can generate unmarked text indefinitely. No regulation reaches a model running on someone's own hardware.
What we do, and what we do not.
We do not read Claude's watermark. Nobody outside Anthropic can right now, and we would rather say that than sell a check we cannot run. When a way to verify the mark is published, we will add it as one more signal alongside the others.
What we do is the part the watermark was never going to cover. We scan a draft against a real detector, show you which passages carry the strongest signals and why, and keep a timestamped history as you revise, so you finish with a record of how the writing actually developed.
That record is the thing worth having. It answers the question a watermark raises but cannot settle.
How to prove you wrote it yourself goes deeper on building that record before anyone asks for it.
Questions
Can I check whether text carries Claude's watermark?+
Not yet, and not with any tool that exists publicly as of August 12, 2026. Anthropic has said it intends to publish a way for people to check, but nothing has shipped. Any site claiming today that it reads Claude's watermark is guessing, usually by running an ordinary AI detector and relabeling the result.
Does the watermark mean AI wrote my document?+
No. It means a Claude model produced some of the words. If you wrote a draft and used Claude to tighten the phrasing, the tightened sentences can carry the mark while every idea, source, and argument remains yours. The mark records processing, not authorship, and Anthropic's own documentation says it should not be treated as proof of who wrote something.
Can I turn the watermark off?+
There is no published opt-out, including for paid or API access. The marking is applied by the model as it writes rather than added afterward, so it is not a setting exposed to the people using it.
Will other AI companies start marking text too?+
Some pressure exists, since the same European rules apply to everyone serving the EU. But OpenAI has deliberately declined to watermark text, Google's text marking cannot be read without its own key, and open models cannot be marked at all. Expect partial coverage for a long time rather than a single universal system.
Does a missing watermark prove a person wrote it?+
No, and this is the more common mistake. Text from an older model, from a provider that does not mark, from an open model, or text that was heavily rewritten or translated all come back with nothing to find. Absence of a mark is not evidence of anything.
Can someone fake a watermark on my writing?+
Researchers have shown that watermark signals can be forged onto text the model never produced. That is one more reason a mark on its own is a starting point for a conversation rather than a finding, and why a record of how a document was written matters more than any single check.
Keep a record, not just a score.
Scan a draft, see what gets flagged and why, then revise and rescan with every version saved.
Start free