What Claude's watermark means for your content

Since August 2026, text from the newest Claude models carries an invisible watermark. If you draft blogs, newsletters or LinkedIn posts with Claude, that mark travels with whatever you publish. So it pays to understand what it does before someone on LinkedIn tells you your blog is about to get flagged.

TL;DR

Claude models released from August 2026 hide a watermark in the text they write. Nothing is added to your copy, there are no hidden characters, and no user ID sits inside it, so it cannot be traced back to you or your business. It answers one question, whether Claude was probably involved, and it goes thin on facts, code, and light proofreading. Detection sits in a private preview for regulators and similar bodies, so nobody can paste your blog into a checker today. Your published work is safe, your phrasing is what readers will judge.

How the watermark works

Claude writes one token at a time. At each step it holds a shortlist of sensible next words and picks one. Plenty of those picks are coin flips: in "the weather today was cold and...", overcast and grey both work, and your reader will never know which one got chosen. Normally a random number settles it.

Watermarking changes where that randomness comes from. Instead of an arbitrary random number generator, Claude uses a secret key plus the handful of words that came just before, so the picks stay unpredictable while following a pattern the key can recognise later. Repeat that across a few thousand word choices and the finished text carries a statistical signature.

Nothing gets added to your text, so there are no hidden characters, no invisible unicode, and no extra tokens on your bill. The method is a version of SynthID-Text, which Google DeepMind published in Nature in 2024 and released a reference implementation for, so this is a known quantity rather than a black box Anthropic invented last month.

The pi analogy

Anthropic's explainer reaches for Monopoly, which is the clearest version I have seen. Players normally roll dice. Swap the dice for a book of the digits of pi, start from an arbitrary position, and everyone reads off the next digit as their roll. The game plays identically, nobody at the table notices, and the moves stay unpredictable. Keep the record of every move though, and anyone who knows pi can work out afterwards that the dice never came out of the box.

Claude's text works the same way. The reading experience does not change, the record can still be checked.

What it can and cannot tell you

The watermark answers one question, with a probability attached: was Claude likely involved in producing this text?

It carries nothing about you, your company, or your chat history. There is no user ID in there, no way to trace a passage back to a person, and no key that unlocks anything about who typed the prompt. It cannot confirm a piece of text was written by a human, and it cannot spot other AI systems, because their keys are different. One more limit: a positive result cannot separate "Claude wrote this" from "Claude edited this heavily".

Where the mark goes thin

The watermark only lives in choices that were free, which gives it predictable blind spots.

Facts leave no room, so a sentence like "Isaac Newton's most famous work was called Principia..." has exactly one correct continuation and the watermark has nothing to sit on. Code is much the same, since it usually has to be exact, though comments inside the code are fair game because their wording is arbitrary. Proofreading barely registers either, because when you hand Claude your own draft and ask for grammar fixes, nearly every word stays yours. Short passages are difficult too, as detection needs volume and confidence climbs with length.

Translation sits at the other end, where Claude picks every single word, so a translated piece is fully marked.

What this means if you publish AI-assisted content

Light editing will not strip the mark. Replacing every word will, at which point you have written the thing yourself anyway.

Detection stays closed for now. Anthropic runs a detection API in private preview for organisations with a legal reason to check: regulators, law enforcement, newsrooms, fact-checkers, researchers, education bodies, EU civil society groups, and enterprises with their own compliance duties. Your competitor cannot paste your blog into a box and get a verdict, and Anthropic's support documentation is where any change to that will show up first.

The reason any of this exists is the EU AI Act. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 other signatories, and it applies the watermark globally because scoping it by region is not workable yet. The obligation underneath the code is Article 50, whose transparency duties became enforceable on 2 August 2026. Files get separate treatment, so a .png, .jpg, or .svg produced by Claude carries a signed C2PA credential in its metadata, the same standard camera makers use, which a screenshot strips out.

Anthropic named the tell we already banned

Buried in the same write-up is a note about how third-party AI detectors work when they do not hold the key. Those tools read the phrasing habits models fall into, and the example Anthropic reaches for is the "this isn't X, it's Y" construction. Its been on our banned list since day one.

A watermark tells a regulator that Claude was in the room. The phrasing tells your reader, by paragraph two, that nobody sat with the piece before it went out. Sort the second problem and the first one stops bothering you.

FAQ

Can anyone check whether Claude wrote my blog?

Not today. Anthropic keeps detection in a private preview for regulators, law enforcement, media, fact-checkers, researchers, educational bodies, EU civil society groups, and enterprises with their own compliance duties, and says it plans to widen access over time.

Does the watermark identify me or my company?

No. It holds nothing about the user, the organisation, or the conversation, and the key cannot recover any of it.

Does watermarking make Claude slower or worse?

No. It produces no extra tokens, so the model costs the same to serve, and when DeepMind ran the technique on a slice of live Gemini traffic the thumbs-up and thumbs-down ratings showed no statistically significant difference.

Will editing my draft remove it?

Light editing leaves it in place. A rewrite that replaces every word removes it, though by then the piece is yours.

What happens with images and other files?

Files carry a signed C2PA credential in their metadata instead of a watermark, and a screenshot or a format conversion drops it.

Which Claude models are marked?

Models launched on or after 2 August 2026 support marking at launch, and Anthropic says it is adding marking to older models over the following months.