Guides
What Is The Claude Watermark?
What is the Claude watermark, in one sentence: an invisible provenance mark that Anthropic adds to Claude's output so machines can later tell the content passed through Claude.
It arrived on 2 August 2026, applies worldwide, and cannot be turned off.
An invisible, mandatory provenance mark on Claude output. Live since 2 August 2026.
Runs in your browser. No signup, no upload, nothing stored.
The facts, compressed
Everything most people need is in this list:
- Live since 2 August 2026 on models launched from that date
- Covers Claude, the API, Claude Code, Claude Cowork and Claude Tag
- Applies worldwide, not only in the EU
- Text gets an invisible statistical mark; .svg, .png and .jpg get C2PA metadata
- There is no opt-out at any tier
- No public detector has been released
What it proves — and does not
Anthropic is unusually careful here, and the nuance matters far more than the headline does.
A detected mark means content may have been processed by Claude. Anthropic describes this as not fully conclusive.
It does not establish that Claude authored the content, and it does not establish that the content is unchanged since it was generated.
Why 'processed' is the load-bearing word
Ask Claude to proofread an essay you wrote yourself and the text that comes back is marked.
Nothing about that mark distinguishes it from text Claude drafted from nothing. This is the source of most of the anger about the rollout, and it is a genuine structural limitation rather than a misunderstanding on anyone's part.
Who it actually affects
Anyone publishing Claude-assisted text where AI involvement carries a cost — editorial policies, academic rules, client contracts, platform terms.
Anthropic advises enterprise and API customers to independently assess what Article 50 requires of them as deployers, which is a pointed way of saying the obligation does not stop with Anthropic.
What it is not
It is not a visible logo, a footer, or a line of attribution. Nothing appears in your document.
It is not DRM, and it does not restrict what you can do with the output. It is a provenance record, which is a weaker and narrower thing than most first reactions assume.
What changed on 2 August
Before that date, Claude output carried no machine-readable provenance signal at all, and detection meant guessing from style.
After it, the signal exists in everything a supported model produces — but the tool to read it does not. That gap is the strangest feature of the current moment, and it will not last indefinitely.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.