Claude now watermarks everything it writes — including this post
On 14 August, Anthropic explained how Claude's text watermarking works — an invisible signature woven into the words themselves, on everything the model writes. It has been switched on since 2 August for new models, it covers the API as well as the chat apps, and there is no opt-out. That includes this post, which is drafted by a Claude agent before a human signs it off.
If you use AI to write quotes, tenders, product copy or client emails, this is the first change that makes "was this written by AI?" a question with an actual answer.
What the watermark is
When a model writes, it picks each word from a handful of plausible candidates, and which one it lands on is normally down to a random number. Watermarking replaces that randomness with something reproducible: a secret key plus the words already written decide the choice. The text reads exactly the same — nothing is inserted, nothing is spelled oddly — but anyone holding the key can test a passage and see whether the run of word choices matches what the key would have produced.
A few details worth having straight:
- It carries nothing about you. The mark says "a Claude model may have touched this". It holds no user, no account, no organisation, no chat history. This is not tracking.
- It applies across the board — the Claude apps, Claude Code, and the API that sits behind other companies' products. Building on the API is not a way around it.
- Images get a different treatment. PNG, JPG and SVG files come with signed C2PA provenance metadata rather than a woven mark — and metadata is routinely stripped when a file is uploaded to a social platform or run through an image editor.
- Nobody can check yet. The detection tooling is described as forthcoming. Today the mark is being laid down; the reading comes later.
The driver is regulatory. Anthropic is one of roughly 190 signatories to the EU Code of Practice on transparency of AI-generated content, signed in July, which sits under the EU AI Act. Australia has no equivalent obligation — but the model is the same model wherever you use it, so Victorian businesses get the marking regardless.
Where it stops working
The limits are as interesting as the capability, and Anthropic is upfront about them. A short passage doesn't carry enough word choices to be conclusive. Neither does a sentence with only one sensible way to say it — a line of dates and figures leaves the model no room to choose, so no room to sign. Code that has to compile exactly the same way carries very little mark. And a genuine rewrite — not a tidy-up, a rewrite — removes it.
So the mark is evidence, not proof. Its presence suggests a Claude model was involved. Its absence proves nothing at all: the text might be human, might be another vendor's model, might be Claude output that someone reworked properly.
What we'd tell a client
Nothing here requires you to stop using AI. It changes what you should assume about the paper trail.
- Decide your disclosure position now, before someone asks. One line in your terms — "we use AI tools to draft, and a person reviews everything before it goes out" — is defensible, unremarkable, and far better than being asked after the fact. The businesses that will look bad are the ones that claimed hand-crafted.
- Ask the same question of what you buy. If an agency, copywriter or offshore team invoices you for original writing, the marking regime cuts both ways. Put an AI-use clause in the contract rather than trying to detect it later.
- Don't launder it. Running output through a paraphraser to scrub the mark is the one workflow this genuinely breaks — and it produces worse writing than either the model or you would have on your own. Editing for accuracy and voice is the useful version, and it leaves the mark largely intact. That's the right way round.
- Regulated or evidentiary writing needs a human in the seat. Anything that could end up in front of a regulator, a court or an insurer should be written and checked by a named person, watermark or no watermark.
Our own position, since we're in the same boat: the drafting here is done by an agent, the facts are checked and the judgement is ours, and we'd rather say so than pretend otherwise. That's the same standard we set up for clients when we build AI into a business process, and it's the governance session in our AI training — not because a watermark is scary, but because "who is accountable for this text" was always the real question. The watermark just made it checkable.