Does Claude watermark its text?

What Anthropic actually did

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act became enforceable. Anthropic responded by marking Claude output, and did it globally rather than only in Europe. According to Anthropic's own help centre, the marks cover output from supported models across the API, Claude, Claude Code, Claude Cowork, and Claude Tag — “and wherever Claude is offered, worldwide.”

Models launched in the EU on or after 2 August 2026 support marking at launch. Older models fall under a transition period, which means not everything Claude has ever written carries a mark.

What kind of watermark is it?

This is the part that trips people up. It is not an invisible character hidden between words. It is a statistical pattern woven into the model's word and token choices as the text is generated — a subtle bias in how it picks among near-equivalent options.

That difference has real consequences. Because the mark lives in the words themselves rather than in a separate character or a metadata field, it survives copy-and-paste, and Anthropic says it “may persist through some editing.” A metadata credential dies the moment you paste text into a new document. This does not.

Why can't anyone detect it?

Because the instructions have not been released. Anthropic has said it will “share details on detection mechanisms in forthcoming technical documentation,” and an engineer has indicated a detection API is coming. Until that lands, there is no way for an outside developer to build Claude watermark detection into anything, including this site.

We would rather say that plainly than pretend. Our file audit reads the signals that genuinely can be read today — cryptographic content credentials, metadata, hidden characters, file integrity — and the signals table lists what remains unreadable, including this. When Anthropic publishes, we will add it, and not before.

The part that matters most

Suppose detection ships tomorrow and you run a document through it and it comes back marked. What have you learned?

Less than you think. Anthropic is unusually direct about this: “Detecting a Claude mark tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content.”

Read that word: processed. Someone who wrote an essay themselves and asked Claude to fix the grammar carries the same mark as someone who had Claude write the whole thing. A translation carries it. A tidied-up email carries it. The mark cannot distinguish between them, because it is applied to whatever text passes through the model, regardless of who thought of it.

Anyone who treats a detected mark as proof that a person did not write something is making an accusation the evidence does not support. That is the single most likely way this technology gets misused, and it will fall hardest on students and non-native speakers.

And absence proves nothing either

The reverse error is just as common. Anthropic again: “Lack of a detected mark doesn't mean the content wasn't AI-generated or processed.”

There are plenty of ordinary reasons a mark would be missing:

• The text came from an older Claude model that predates marking.
• It came from a different provider entirely.
• The passage is too short — a sentence or two does not contain enough material for a statistical signal to be reliable.
• It was heavily rewritten after generation.

So a negative result is not a clean bill of health, and a positive result is not a confession. Treat either as one input to a human judgement, never as the judgement itself.

What you can check right now

While text watermark detection is unavailable to everyone, file provenance is not. If you are dealing with images, PDFs, or video, C2PA content credentials are readable today and are genuine cryptographic evidence rather than a probability. The difference between the two mechanisms is worth understanding before you rely on either.