Claude now watermarks everything it writes. It proves less than you think.

Anthropic began marking Claude’s output this month. Text gets an invisible watermark woven into the words themselves; generated image files get signed provenance metadata in the C2PA format. It applies to models launched on or after 2 August 2026, across the API, Claude Code and the rest, and — this is the part people missed — worldwide, not just in the EU, even though EU law is what prompted it.

The coverage I’ve read frames this as a detector: at last, a way to catch AI writing. Schools can check homework, employers can check cover letters, you can check whether your copywriter actually wrote it.

Anthropic’s own documentation says otherwise, in plain language, and it is worth reading before anybody starts accusing anybody.

The mark says “passed through”, not “written by”

“Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.”

Read that again with your own working week in mind.

You write a difficult email to a customer yourself, then ask an AI to tidy the tone. Marked. You write a property description and have it translated. Marked. You dictate your own thoughts and ask for them to be tightened into three paragraphs. Marked. In every case the thinking was yours and the mark is there anyway.

A watermark is a record of a tool having touched the text. It is closer to a photocopier’s serial number than to a signature.

And it disappears from things it did write

The failure runs the other way too. Anthropic notes that marked content “may be modified, excerpted, or combined with other material”, and may carry no detectable mark where “the text has been heavily edited, paraphrased, translated, or mixed into other writing”.

So: a mark does not prove a machine wrote it, and no mark does not prove a person did. Anyone actually trying to pass off AI work as their own will edit it, because they always have. The people who will get flagged are the honest ones who ran their own words through a spell-check with ambitions.

That is not a criticism of the watermark. It is doing exactly what it says on the tin — provenance, not authorship. The problem is entirely in how it is being described.

What this changes for a small business

Practically, very little — if you are already open about it. If your website says you use AI to draft listings and a human checks them, a watermark confirms your own account of yourself. There is nothing to catch.

More, if you have been quiet about it. Not because you will be caught — see above, detection is weak in both directions — but because the ground is moving. Marking is becoming a default across the industry, and “we never said” ages badly next to “we said from the start”. We wrote about the cheap version of this in how to tell customers you use AI, and the legal backdrop in what the EU AI Act disclosure rules ask of you.

And be careful what you do with a detection. If you are an employer or a school looking at a marked document, what you have learned is that the text met an AI at some point. Acting on that as though it were proof of cheating is a good way to be badly wrong about a real person.

What argues the other way

The watermark is genuinely useful for the thing it was built for. Provenance at scale — being able to ask “has this document been near a model at all?” across millions of files — is a real capability, and C2PA signing on images is a stronger signal than anything we had before. For fraud investigation, or tracing where a synthetic image entered circulation, that matters.

It is also an early, voluntary move on a rule that has barely started biting, which is more than most of the industry has done. Anthropic is being straighter about the limits than the coverage has been.

And a fair point against my own argument: over time, if marking becomes universal and unedited AI output becomes easy to spot, the norm may shift on its own — not because detection works perfectly, but because assuming it might is enough to change behaviour. That would be a real effect, just not the one being advertised.

The short version

A watermark tells you a machine handled the words. It does not tell you a machine thought of them, and its absence tells you almost nothing. The reliable way to be trusted about AI has not changed and is not technical: say what you use, and stand behind what you publish.

Source: all quotations are from Anthropic’s support article “How Claude marks AI-generated content”, read directly. Dates and scope (models launched on or after 2 August 2026; worldwide application; API, Claude, Claude Code, Claude Cowork and Claude Tag) are from the same page. Anthropic also notes a transition period for models launched before that date.