Claude Now Marks What It Writes. Here's What That Actually Proves.

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and has begun putting it into practice. In plain terms: Claude is starting to mark what it produces, so that software can later tell the content passed through Claude.

The detail that matters most for anyone reading this from Kuala Lumpur or Singapore is buried in the middle of Anthropic's announcement: marking applies wherever Claude is offered, worldwide. This is a European law, but you do not have to be in Europe to be affected by it. If you use Claude, this applies to your output too.

Here is what is actually happening, and — more usefully — what a mark does and does not prove.

The two mechanisms

Invisible watermarks in text. When a supported Claude model generates text, it weaves an imperceptible watermark into the text itself. You will not see it. Anthropic states it does not change the meaning, quality or readability of the response.

Because the mark lives in the text, it travels when the text is copied and pasted elsewhere, and it may survive some editing. It is applied at the model level, which means it is present regardless of which Claude surface produced it.

Signed provenance metadata on files. When Claude generates a supported file — .svg, .png, .jpg — it attaches signed provenance metadata following the C2PA open standard, the same industry standard used elsewhere for content provenance. That metadata signals the file was processed by Claude, and allows detection of whether it has since been tampered with.

Where it applies

• Models: those launched on or after 2 August 2026 support marking at launch. Earlier models fall under a transition period and Anthropic says support is being added.
• Products: Claude Platform (the API), Claude, Claude Code, Claude Cowork and Claude Tag.
• Cloud partners: embedded text watermarks apply when supported models are accessed through AWS, Google Cloud or Microsoft Foundry. Signed provenance metadata may not be supported on every platform.
• Regions: worldwide.

Anthropic also says it will enable users and third parties to detect the marks, with technical documentation to follow.

If you just use Claude, almost nothing changes

Practically speaking: nothing about your day gets worse. The watermark is imperceptible and does not affect quality. You do not need to do anything, turn anything on, or disclose anything you were not already disclosing.

The one thing worth internalising is that the output is no longer anonymous by default. Paste a Claude-written paragraph into an email, a proposal, a WhatsApp message or a Word document, and the mark travels with it. That is not a trap — it is simply the new default, and it is the same direction the whole industry is moving.

If you produce content, read this part twice

This is where the practical consequences live, and where the misunderstandings are going to be expensive.

Mistake one: assuming a mark means "AI wrote this"

It does not. Anthropic is explicit that a detected mark is a signal, not a conclusion, and gives the reason plainly: people use Claude to proofread, translate, summarise and convert files. The output can carry a Claude mark even when the ideas, the text and the data came from somewhere else — including from you.

Think about what that means in practice. You write an article yourself. You ask Claude to tighten the grammar. The result may carry a mark. Your work, your thinking, your sentences — marked, because a tool touched it on the way out.

Anyone who treats a detected mark as proof of authorship is going to accuse the wrong people.

Mistake two: assuming no mark means "a human wrote this"

This one is worse, because it is the direction institutions will lean.

Anthropic lists the ways marked content loses its mark: text generated by a model released before marking was supported; text that has been heavily edited, paraphrased, translated or mixed into other writing; passages too short to carry a reliable signal; file metadata stripped by format conversion, re-saving or screenshots; and output from any platform, feature or file type where that marking type is not supported.

Absence of a mark tells you close to nothing.

The uncomfortable inversion

Put those two together and you get a genuinely awkward result that nobody is going to enjoy:

The mark is most likely to survive on content that was copied out and used as-is, and most likely to disappear from content that was heavily rewritten by a person.

So the signal correlates with how little a human intervened — which sounds useful until you realise it also means the careful writer who reworks every paragraph looks identical to someone who wrote it from scratch, while the person who published a raw draft is flagged. That is a reasonable outcome for a transparency signal. It is a terrible basis for an accusation.

The rule for content teams: use marks as provenance context, never as an authorship verdict. If your editorial, HR or academic policy is heading towards "we will run a detector and act on the result", that policy will produce false accusations against real people. Build the policy on disclosure and process instead.

If you build products on Claude

Anthropic's guidance here is short and worth quoting the shape of: if you deploy Claude inside your own product, you should independently assess what Article 50 requires of your own products and services. Anthropic's commitments cover Anthropic's models; they do not discharge your obligations as a provider or deployer.

For Malaysian and Singaporean businesses this is easy to wave away as a European problem. It is not, for two reasons. First, the marks are being applied worldwide regardless of where you operate. Second, if you sell software or services into the EU, the Act's transparency obligations can reach you.

If you are shipping anything customer-facing built on Claude, the sensible move now is to write down what you disclose, to whom, and where — before anyone asks. That is a small piece of work today and a scramble later.

What we would actually do this week

1. Check your disclosure language. If you publish AI-assisted content, say so in a way that is true. This change makes the claim checkable in a way it was not before.
2. Do not build detection into policy. Not for hiring, not for coursework, not for editorial review. The false-negative and false-positive paths are both wide open, and Anthropic says so itself.
3. Know which surfaces you use. Marking applies across the API, Claude, Claude Code, Cowork and Tag, and through the major cloud platforms — but file provenance support varies by platform.
4. If you ship a product on Claude, assess Article 50 for yourself. Anthropic's compliance is not your compliance.
5. Expect the detection tooling to arrive. Anthropic has said documentation is coming. Decide now how you will use it, so the answer is not invented under pressure.

The honest summary

This is a good change, and a modest one. It gives people some context about where content came from, which is worth having as AI-written text becomes ordinary.

What it does not do is settle arguments about authorship. A mark says content may have passed through Claude — not that Claude thought of it, and not that a person did not. The absence of a mark says almost nothing at all.

Treat it as provenance context. Anyone selling you a detector as a lie detector is overselling it, and Anthropic's own documentation is the best evidence of that.

Sources and standing. Details here are taken from Anthropic's Claude Help Center article How Claude marks AI-generated content, read on 11 August 2026, and Anthropic notes it will update that article as more technical guidance becomes available — check it before acting on anything time-sensitive. Anchor Sprint is a member of the Anthropic Claude Partner Network; we are not an official, exclusive or endorsed representative of Anthropic, and nothing here is legal advice. Claude and Anthropic are trademarks of Anthropic, PBC. For obligations specific to your business, take advice from your own counsel.